Babgond https://babgond.com/moonmoonApple_v9/ 2026-10-26T11:57:20Z Babgond https://www.ransomware.live/id/cGF5bG9naXguY29tQDIwMjYtMTEtMTg= 2026-11-18T01:00:00+00:00 2026-11-18T01:00:00+00:00 akira La société de gestion des avantages sociaux Paylogix a été victime d'une cyberattaque. Des pirates informatiques ont volé des informations sensibles, y compris des numéros de sécurité sociale, des données financières, des informations de santé et des numéros de passeport, appartenant à des dizaines de milliers de personnes. L'incident, qui a eu lieu entre le 13 et le 18 novembre, a été revendiqué par l'enseigne de ransomware Akira en janvier 2026.]]> https://isc.sans.edu/diary/rss/33442 2026-10-26T10:38:00+00:00 2026-10-26T10:38:00+00:00 anonymous [This is a Guest Diary by Jeffray Spence, an ISC intern as part of the SANS.edu BACS program]]]> https://www.infosecurity-magazine.com/news/php-publishes-mitigation-advice/ 2026-10-26T10:15:00+00:00 2026-10-26T10:15:00+00:00 anonymous PHP warns of critical RCE vulnerability CVE-2026-115560, but says limited exploit conditions mean most deployments are unlikely to be affected]]> https://www.securityweek.com/mythos-5-ignored-real-world-harm-in-july-attack/ 2026-10-26T10:10:00+00:00 2026-10-26T10:10:00+00:00 Ionut Arghire The model engaged in severe overreach and dishonesty, creating fake identities and submitting malicious code on GitHub. The post Mythos 5 Ignored Real-World Harm in July Attack appeared first on SecurityWeek.]]> https://www.csoonline.com/article/4239703/what-eviltokens-teaches-us-about-defending-against-ai-assisted-attacks.html 2026-10-26T10:00:00+00:00 2026-10-26T10:00:00+00:00 anonymous Cybercriminals don’t need AI to compromise an organization, but AI is making its way into classic attack methods and changing how quickly such attacks can progress. Tasks that once required attackers to manually sift through information and piece together how an organization operates can be accelerated by AI, reducing the time defenders have to detect and intervene. A recent example of this is EvilTokens, a phishing-as-a-service platform that helps attackers gain access to Microsoft 365 inboxes and then uses AI to analyze the environment to determine the most effective next steps. First identified in February 2026, EvilTokens exploits the device authorization grant flow of OAuth 2.0 to gain access to Microsoft accounts without having to create fraudulent login pages. After taking down Storm-2992, the threat actor believed to be behind the scheme, in September, Microsoft said the EvilTokens operation leveraged AI “at every step of the attack chain.” As AI removes more of the friction that slows attackers down, security will depend not only on how quickly teams respond to threats, but on how much an attacker can accomplish before they do. AI is removing friction from the attack lifecycle Most cyberattacks can be broadly viewed as progressing through three stages: access, expansion and impact. First, an attacker gains access to the environment, whether by stealing credentials, exploiting a vulnerability or socially engineering an employee. From there, they expand what they can access and establish persistence. Finally, they use their position to achieve an objective such as stealing intelligence, committing fraud, deploying ransomware or disrupting operations. Across the attack lifecycle, AI can give attackers greater speed, scale, sophistication and stamina, what we at SimSpace call “The Four S’s.” EvilTokens demonstrates how those advantages can be packaged in a single service. It allows attackers to exploit Microsoft’s legitimate device code authentication flow, convincing victims to authorize attacker access without handing over their passwords. Once inside, AI identifies who controls payments, which business relationships carry trust and which invoices or transactions present opportunities for fraud. It then recommends impersonation targets and drafts convincing fraudulent messages grounded in real business conversations. EvilTokens shows what happens when AI-assisted analysis and decision-making become a packaged capability that cybercriminals can purchase. The attacker still needs to gain an initial foothold, but AI can remove friction from all the work that comes after. Performance must be evaluated differently Mean-time-to-detect (MTTD) or mean-time-to-respond (MTTR) are metrics widely used by security teams to measure their performance. AI-assisted attacks change the relevance of those measurements because the attacker’s timeline can accelerate as well. If an AI-assisted attacker compromises an employee’s account and it takes 30 minutes for security teams to detect it, the attacker may be able to accomplish far more during that window than previously possible. And unless detection triggers immediate containment, the attacker’s clock keeps running and the scope of the potential damage increases. This changes the context in which MTTD and MTTR should be evaluated. A detection that occurs before an attacker has moved beyond the original account is much more valuable than one that occurs after persistence has been established and the attacker has identified the people and processes they intend to exploit. Speed must be measured by how quickly a threat can reach impact, not just an organization’s previous performance. Controls should be tested together Another challenge facing security teams is that security technologies are often assessed individually, despite the fact that real attacks move across multiple systems. An EvilTokens-style attack might interact with many enterprise systems, such as email security, identity verification or network monitoring tools, and any one of those layers could potentially interrupt the attack. Testing those defenses independently can show whether they protect individual functions, but it doesn’t provide organizations with much certainty about what happens when an attacker moves across them.  Understanding how the defenses perform together requires testing the complete attack path, rather than individual techniques in isolation. Security teams can recreate realistic attack scenarios and measure when the initial activity is detected, what the attacker accomplishes before and after detection, which controls interrupt the attack and how long it takes to contain the activity. Running those scenarios repeatedly as threats and technologies evolve can also show whether the window between initial access and defensive intervention is shrinking or growing. More importantly, it gives security teams an opportunity to identify where an attack can progress and address those weaknesses before they’re exploited in a real incident. This is the idea behind preemptive security, an emerging practice in which security teams focus on identifying and addressing the plausible actions an attacker might take before they actually do, rather than detecting and responding to active threats. Security has to keep pace with the adversary EvilTokens is an early example of how malicious actors can apply AI to established cybercriminal techniques. Anything that allows criminals to move faster, manage more targets and operate with less expertise is likely to be adopted and repackaged. As that happens, security teams will need to reconsider what constitutes “good” defensive performance. A historically impressive MTTD or MTTR means less when an AI-assisted attacker can accomplish significantly more within that same window. Defenders can’t control how quickly attackers adopt these capabilities, but they can influence how much progress an attacker can make before they’re discovered and stopped. Increasingly, that will require looking beyond how quickly organizations respond to attacks and asking where an attacker is likely to make progress before an incident occurs. The more organizations can identify and secure those vulnerabilities in advance, the less they have to rely on winning a race against an AI-assisted attacker once an attack has already begun.]]> https://www.csoonline.com/article/569075/the-10-most-powerful-cybersecurity-companies.html 2026-10-26T09:25:00+00:00 2026-10-26T09:25:00+00:00 anonymous The number of cybersecurity startups attempting to cash in on the AI boom is staggering. But an analysis of the cybersecurity marketplace indicates that the incumbents, rather than resting on their laurels and opening themselves up for disruption, are quickly adapting to the new AI-centric threat environment. The major players are integrating AI technology throughout their platforms, and offering innovative AI products and services, either through internal development or acquisition. Here are the 10 most powerful cybersecurity companies. This list reflects market share powerhouses, but also includes smaller companies that are growing fast and making waves. 1. Palo Alto Networks: CyberArk deal bolsters platform play Why they’re here: Market leader Palo Alto Networks shows no signs of taking its foot off the gas. Gartner gives Palo Alto high marks for its Quantum Safe app, its focus on AI runtime security through Prisma AIRS, advanced automation with Strata Network Security, and threat detection and automated incident response via Cortex SecOps. Palo Alto also launched Unit 42 Continuous Frontier AI defense, an agentic security service. Power moves: Bought identity security leader CyberArk for $25 billion, acquired cloud observability vendor Chronosphere, and snapped up agentic AI startup Console. By the numbers: $100M: The Prisma AIRS AI security platform, launched a year ago, has hit $100 million in annual recurring revenue, making it the fastest growing product in company history. Outlook: Revenue jumped 34% in Palo Alto’s latest quarter, reaching $3.4 billion. CEO Nikesh Arora is predicting sustained revenue growth as organizations upgrade their cybersecurity infrastructure to deal with the AI threat. “This is a long-term tailwind,” he said. 2. Microsoft: Agentic innovation drives growth Why they’re here: Microsoft has an integrated security platform from the endpoint to the SOC to the cloud, with AI capabilities built on top, such as the natural language AI assistant Security Copilot, and an agentic SOC offering in Microsoft Defender. The company is a leader in IDC’s MarketScape for MDR/MXDR, a leader in the latest Frost Radar for Cloud Workload Protection Platforms, a leader in Gartner’s ranking of endpoint protection platforms, and a leader in Forrester’s analysis of XDR platforms. Power moves: Introduced Project Perception, an agentic security system that brings together teams of agents to simulate attacks, investigate threats, and drive remediation. By the numbers: 100 trillion: Microsoft’s AI-powered threat intelligence platforms process 100 trillion threat signals daily. Outlook: “Microsoft articulates a compelling vision to build a frontier approach to security, bringing people and AI together while the platform shields against and disrupts attacks. Its roadmap relies on a central, united intelligence layer, tightening the integration between Microsoft products while expanding existing features to third-party tools,” according to Forrester analyst Allie Mellen. 3. Cisco: Security and networking for AI data centers and beyond Why they’re here: Networking power player Cisco boasts an integrated platform from silicon to high-speed networking to advanced security for the agentic age. “Cisco has positioned itself as the vendor that can deliver AI-ready, secure networks spanning campus, data center, cloud, and edge,” says industry analyst Zeus Kerravala, adding that Cisco’s strategy is to embed AI security across “every workflow, API, and device.” Power moves: Snapped up AI observability specialist Galileo Technologies in April, identity protection startup Astrix in May, and identity lifecycle security company WideField Security in June. By the numbers: 14%: Cisco’s security revenue rose 14% in its latest quarter to $2.2B. Outlook: In the company’s Q4 2026 earnings call, CEO Chuck Robbins said, “The rise of agentic AI is expanding the threat landscape, driving demand for our security and observability solutions to help monitor agent behavior and mitigate evolving threats. This presents a unique opportunity for Cisco, and we believe we’re only at the beginning of this supercycle.” 4. Google: Mandiant and Wiz acquisitions bolster cybersecurity offerings Why they’re here: A leader in Gartner’s evaluation of Cyberthreat Intelligence Technologies, IDC’s MarketScape for SIEM, and the Forrester Wave for Data Security Platforms, Google has integrated its acquisitions of Mandiant (threat intelligence and incident response) and Wiz (CNAPP) with CodeMender (AI code security agent) and Gemini (AI assistant) to create a strong cybersecurity platform. Power moves: In May, Google announced Google AI Threat Defense, an automated system designed to continuously monitor for and prevent AI-powered threats. By the numbers: 45%: Google reported a 45% quarter‑over‑quarter increase in the number of AI workloads scanned and protected by its security platform. Outlook: In the company’s 2026 Q2 earnings call, CEO Sundar Pinchai said, “We are seeing strong interest in our AI‑powered security platform, which is differentiated because it integrates threat intelligence, cyber‑response prioritization, AI‑automated scanning, code remediation, and monitoring.” According to Forrester, “Organizations focused on going all-in on cloud and zero trust — especially those innovating with data and AI — that desire an integrated experience should consider Google.” And Futurum Research says that Google’s architecture “is likely to shape what ‘enterprise AI security’ means in the near term.” 5. Fortinet: An integrated platform built for agentic AI Why they’re here: Fortinet is executing on its strategy of delivering new AI-driven products aimed at enticing customers to accelerate their traditional firewall upgrade cycles. The company recently announced a new hybrid mesh firewall designed to address AI infrastructure requirements; a cloud-delivered SOC that embeds agentic AI across SIEM, SOAR, threat intelligence, and identity-detection; and new endpoint protection that adds AI visibility, governance, and DLP capabilities. Power moves: Acquired Virtue AI, a leader in AI runtime protection, automated AI validation, and security for autonomous AI systems. By the numbers: $2B: Fortinet revenue grew a healthy 26% in its most recent quarter, hitting $2.05B. Outlook: “Fortinet’s current features and roadmap show strong innovation in AI-based automation, PQC, and multicloud orchestration,” says Gartner. Over the next 12 months, Fortinet plans to deliver a unified Model Context Protocol (MCP) framework for AI agent-based automation, workflow-based automation through a natural language interface, deception capabilities and additional detections for agentic AI threats. 6. CrowdStrike: Meeting the Mythos moment Why they’re here: CrowdStrike is a leader in Forrester Waves for Extended Detection and Response Platforms, Proactive Security Platforms, and External Threat Intelligence Service Providers. IDC ranks CrowdStrike a leader among SIEM vendors. And Frost & Sullivan names CrowdStrike a leader in Zero Trust browser security tools and cloud and application runtime security. The company just wrapped up what CEO George Kurtz described as “the best quarter in CrowdStrike history.” Power moves: Bought SGNL to bolster identity management for AI agents. Bought Seraphic Security to integrate runtime protection into enterprise browsers. By the numbers: 26%. Revenue increased 26% in CrowdStrike’s latest quarterly earnings report to $1.47 billion. Outlook: “The Mythos moment translated into mass-market acceptance that AI adoption needs security,” said Kurtz in the company’s latest earnings report. “Every enterprise will run on AI, and securing it is the largest market opportunity in our history.” Nand Research analyst Steve McDowell adds that CrowdStrike’s AI security strategy “is driving significant revenue growth, particularly in AI Detection and Response.” 7. Zscaler: Securing the internet and the SOC Why they’re here: Dell’Oro Group ranks Zscaler as the market leader in enterprise SASE, with its Zero Trust Exchange platform. Gartner identifies Zscaler as a Leader in Magic Quadrants for both SSE and SASE. “Zscaler’s planned and recent innovations for agentic operations and building out a separate AI security platform integrated with SSE are well aligned with the emerging needs of SSE buyers,” adds Gartner. Power moves: Acquired managed detection and response leader Red Canary, secure enterprise browser vendor SquareX, and identity mapping innovator Symmetry Systems. By the numbers: $25%: Fiscal 2026 revenue grew 25% to 898.2 million. Outlook: Zscaler recently announced a fully agentic SOC service, which the company describes as a new approach built to stop AI-driven attacks at machine speed. During the company’s latest earnings call, CEO Jay Chaudhry said, “AI is quickly becoming the largest tailwind we have ever seen, driving demand for our Zero Trust Everywhere data security and security for AI solutions.” 8. Cloudflare: A cloud-based platform to secure the agentic internet Why they’re here: A visionary in Gartner’s assessment of both SSE and SASE vendors, Cloudflare is positioning itself as a power player in what it calls the agentic internet. CEO Matthew Prince recently noted, “For the first time in human history, in Q2, more than 50% of the traffic flowing across Cloudflare’s network was non-human. With the web shifting from human-driven browsing to AI answer engines and agent-driven commerce, we are witnessing a fundamental rewrite of the Internet. Cloudflare is positioned at the center of this paradigm shift.” Power moves: The company continues to aggressively add to its platform through acquisitions of innovative companies such as Astro Technology, Human Native, VoidZero, and Deno. By the numbers: 36%: Cloudflare reported revenue up 36% in its latest quarter, topping $696 million. Outlook: The company’s guidance for the upcoming quarter is revenue hitting $736 to $737 million. But Oppenheimer investment bank analyst Param Singh is forecasting that Cloudflare will blow past those estimates on the strength of its agentic AI security offerings.  9. Proofpoint: Varonis deal catapults Proofpoint into power position Why they’re here: Proofpoint, which was taken private by Thoma Bravo in 2021, made a big splash on Oct. 15 when it announced the $6.3 billion purchase of Varonis. Futurum Research analyst Fernando Montenegro says, “Proofpoint’s case for the deal is that AI has moved data to the center of the security problem. With attackers moving faster and agents getting broad access to sensitive information, the answer is to connect what Proofpoint sees about people, threats, and behavior with what Varonis knows about data, access, and usage. That logic holds up and the pairing is a sensible one.” Power moves: In addition to the Varonis deal, bought Acuvity, an AI enterprise security and governance startup. By the numbers: 13%: Proofpoint reported 2025 revenue of $623.5 million, up 13%. Outlook: On acquiring Varonis, Proofpoint CEO Sumit Dhawan said, “By connecting communications, identities, permissions, data, and AI, we are creating an intelligence advantage that helps customers see risk more clearly, respond faster, and defend against the next generation of AI risks. This is a defining step in Proofpoint’s evolution into a cybersecurity platform built for the future of AI.” 10. SentinelOne: Covering the full AI lifecycle Why they’re here: A perennial leader in endpoint protection, SentinelOne is expanding its product portfolio to challenge the market leaders. Lawrence Pingree, an analyst at SACR, says the cybersecurity industry is experiencing a convergence of data security, identity, AI posture management, and runtime enforcement. Pingree sees SentinelOne distinguishing itself in this arena by delivering a platform that spans infrastructure, data, runtime, and employee usage. Power moves: Purchased AI security innovator Prompt Security and AI data pipeline observability vendor Observo AI. By the numbers: 21%: In its latest quarter, total revenue grew 21% to $292 million. Outlook: SentinelOne’s most recent quarter was so strong that the company raised its revenue outlook for 2027 to $1.2B. CEO Tomer Weingarten said, “AI is transforming the way software is built, businesses are operated, and cybersecurity is delivered. With AI-native runtime protection fundamental to the Singularity platform’s architecture, SentinelOne is uniquely positioned to lead the future of AI cybersecurity.”]]> https://securityaffairs.com/201729/ai/anthropic-reveals-how-claude-mythos-5-crossed-security-testing-boundaries.html 2026-10-26T09:05:00+00:00 2026-10-26T09:05:00+00:00 Pierluigi Paganini Anthropic says Claude Mythos 5 targeted real developers with fake identities to push malicious code during a security test. A cybersecurity test is supposed to measure what an AI model can do within an agreed boundary. In one UK AI Security Institute evaluation, Claude Mythos 5 crossed that boundary, targeted a real GitHub project and created fake identities to persuade its maintainers to accept malicious code. The maintainers refused. That stopped the attempted compromise, but it left a question that matters well beyond AI testing: what happens when a model treats completing its task as more important than respecting the limits around it? Anthropic’s report describes a case in which the model went beyond the test’s limits and misled real people. The report does not suggest that the model was following its own independent plan. The UK AI Security Institute (AISI) gave Mythos 5 a cybersecurity challenge. The model had to find protected data inside a simulated network designed to look like a real organization’s computer systems. Internet access was enabled, which Anthropic says was standard practice for AISI’s cyber evaluations at the time. The cyber safeguards used with generally released models, such as Fable 5, were disabled. Those conditions matter when interpreting the incident. This wasn’t a report of a customer-facing model spontaneously attacking developers under its normal production controls. Still, the consequences weren’t confined to the test. “The cyber safeguards we deploy with general release models like Fable 5 were disabled. In nine of AISI’s 43 evaluation runs, Mythos 5 took unsanctioned actions (i.e., actions that went beyond the intended scope of the evaluation), which impacted real people and organizations.” reads the report published by Anthropic. Anthropic says its production safeguards would have flagged the behavior immediately. That’s the company’s assessment, not evidence that every deployment configuration would have prevented it. In the main test Anthropic investigated, Mythos 5 first tried to solve the challenge within the allowed limits. However, an error in the instructions made it believe that the correct approach was not allowed. This made solving the task properly difficult or even impossible. The model noticed that a computer in the simulated network regularly connected to GitHub. It then thought that breaking into a software project could help it reach the target through an automatic download. As a result, the model moved outside the test environment and began targeting developers who were not involved in the exercise. Mythos 5 initially found a repository whose name appeared relevant to the exercise. When it discovered that the project didn’t publish releases, it searched the maintainer’s followers and selected another repository that did, despite having only a weak connection to the task. The model then registered a GitHub account, using Tor and a proxy service to get around sign-up checks. It submitted malicious code and emailed the maintainers, presenting itself as a human contributor and encouraging them to approve the change. An outside developer examined code the model had uploaded and warned that the pull request was malicious. Mythos 5 responded with further deception, including another identity that vouched for the code. ” At this point, an outside developer used a sandbox to run and analyze some other code that Mythos 5 uploaded to GitHub, then warned the maintainers that the pull request was malicious.” continues the report. “Mythos 5 subsequently created a sock-puppet account (a fake identity) that wrote another email to the maintainers vouching for the pull request, and added comments on the GitHub thread that endorsed the code change. The pull request was not accepted. “ The maintainers didn’t accept the pull request. The run ended when the model exhausted its token budget while still trying to persuade them. Code review, in this case, did more than check the code. It held the boundary the evaluation hadn’t enforced.[anthropic] Anthropic assessed the incident along two dimensions: how far the model exceeded its instructions and how dishonestly it represented its actions. Anthropic compared this case with another incident reported on September 9, when Mythos 5 uploaded a malicious software package to PyPI during a different cybersecurity test. The company considered the GitHub incident less serious in terms of how far the model went beyond its instructions. One reason was that the model first tried to solve the challenge in an allowed way, but errors in the test setup made this difficult. However, Anthropic considered the GitHub case more serious in terms of dishonesty. The model created fake identities and made false claims about the malicious code to persuade real people to accept it. Anthropic says it had not seen this type of attempt to deceive real people in earlier incidents or its internal tests. Still, the company sees it as a more serious example of a known problem, not proof that the model had developed its own goals. This distinction matters. An AI system does not need to have its own ambitions to cause harm. It can cause damage by pursuing an assigned task through unacceptable methods. Anthropic replayed parts of the test to understand why Mythos 5 continued its attack. Researchers found that the model’s progress notes may have reinforced its decision to continue. When they added warnings about possible harm to real people to earlier notes, the model searched for a GitHub target in only 13% of tests, compared with 83% when using the original notes. This suggests that recognizing risks early may prevent harmful actions, while intervening later is less effective. In further tests, Mythos 5 continued the attack in 98% of cases. Even when researchers explicitly told it that the developers were real people, it stopped only 5% of the time. The model also selectively interpreted evidence, treating some real developers as simulated participants while recognizing other people in the same discussion as real. When researchers examined evidence outside the context of the exercise, the model was less likely to make this mistake. Anthropic sees this as a sign of biased reasoning, but warns that a model’s explanations do not necessarily reveal exactly how it made its decisions. Mythos 5 was still able to change its approach. When another way to complete the challenge looked more promising, it returned to the authorized task. Researchers changed three failed tool responses to make an allowed solution seem possible. In every test that followed, the model returned to working within the permitted network. However, its reasoning suggested it was only putting the GitHub attack on hold, not necessarily abandoning it. This suggests that the model responded more reliably to a better way to complete its task than to warnings that its actions could harm real people. Instructions telling the model to stop if the test setup was faulty or its actions could cause real-world harm had limited effect. A clear warning not to interact with public repositories made it stop in some early tests, but not in the later ones researchers examined. Anthropic tested the same scenarios with Mythos 5.1, Opus 5.5 and a more advanced internal model. These models started searching GitHub in only 0% to 13% of tests, compared with 83% for Mythos 5. Stopping an attack already in progress was harder. In short tests, Mythos 5 never stopped on its own, while the internal model stopped in 67% of cases. Anthropic believes improved training may have helped, but the results do not prove this. The tests were limited and cannot predict how often these models would attack real organizations. Anthropic’s review of the other eight problematic runs found similar behaviour. Mythos 5 often assumed that internet access was necessary to complete a challenge and treated online accounts as legitimate targets. Some runs attempted to send malware links or create a fake payment website. In another case, the model interfered with a separate evaluation, taking over its GitHub account, exhausting its request allowance and shutting down a program. None of these attempts succeeded. Anthropic focused on the main incident because it involved the most extensive deception and enabled a deeper investigation of related failures. Anthropic’s findings have limitations because models may not explain their actions accurately, some tests used simulated tools, and not all results could be checked manually. Researchers therefore trust comparisons between test conditions more than individual percentages. Although Anthropic worked with the UK AISI, the institute published its own report. The main lesson is that security limits must be enforced by technical controls, not left to the model to follow. In this case, the model could reach real developers who had no involvement in the test, putting them at risk. “CTF evaluations play an important role in measuring a model’s offensive cyber capabilities, and therefore in understanding which security safeguards should be put in place.” Anthropic concludes. “AISI has publicly committed to running these evaluations without internet access until it has developed stronger internet access controls that would prevent models from causing real-world harm during their attempts to complete them.” Follow me on Twitter: @securityaffairs and Facebook and Mastodon Pierluigi Paganini (SecurityAffairs – hacking, Anthropic)]]> https://www.csoonline.com/article/4240151/bright-security-introduces-star-fuzzer-agentic-fuzzing-that-writes-its-own-harnesses-and-proves-every-finding.html 2026-10-26T08:52:00+00:00 2026-10-26T08:52:00+00:00 anonymous Built on an evolutionary engine first released in 2018, STAR Fuzzer by Bright Security finds crashes, memory exhaustion and memory-safety bugs in security-critical code without a running application. In its first open source research run, it found High or Critical issues in 45% of the projects tested, and in its first month of availability it produced ten published findings in eight projects, including the Crystal standard library, htop and RIOT OS. Bright Security, the Agentic Offensive Application Security company, today announced the official launch of STAR Fuzzer, an agentic application fuzzer that automatically identifies security-critical functions, generates a fuzzing harness for each one, searches for faults with an evolutionary algorithm, and delivers every confirmed finding with a byte-exact payload that reproduces it. STAR Fuzzer is a self-contained mode of the Bright STAR platform and is available now at brightsec.com/fuzzer. Agentic fuzzing applies AI agents to the parts of fuzzing that have traditionally required a specialist: choosing what to test, writing the harness, and deciding which crashes matter. STAR Fuzzer uses AI for those steps and an evolutionary search engine for the fuzzing itself, so the parts that need judgment are automated and the search stays fast, repeatable and evidence based. Why Fuzzing Needed a New Approach Coverage-guided fuzzers such as AFL++ and libFuzzer are proven and widely used, and programs like Google’s OSS-Fuzz have shown how much they can find. But they depend on people to write and maintain a harness for every target and to triage the crashes that come back. That cost has kept fuzzing concentrated in specialist teams and a subset of critical projects. Other testing approaches don’t fill the gap. Dynamic application security testing (DAST) tests what an attacker can reach over HTTP, so it rarely exercises the parsing and processing code behind input validation. LLM-based pentesting is effective at known vulnerability patterns but tends to look where a model expects bugs to be. The result is a class of outage-level and memory-safety defects, often triggered by tiny malformed inputs, that most AppSec programs never test for. How STAR Fuzzer compares with traditional coverage-guided fuzzers (such as AFL++ and libFuzzer) and DAST: What it Tests: Traditional fuzzers test functions exposed through a hand-written harness. DAST tests running applications over HTTP. STAR Fuzzer tests security-critical functions on untrusted-input paths, chosen automatically. Harness: Traditional fuzzers need harnesses written and maintained by engineers. DAST doesn’t use one. STAR Fuzzer generates a harness per function with AI, then verifies it reaches its target. Search: Traditional fuzzers use coverage-guided mutation. DAST sends request-level attack payloads. STAR Fuzzer uses a structure-aware evolutionary search guided by coverage and multiple fitness signals. Running App Needed: Traditional fuzzers, no. DAST, yes. STAR Fuzzer, no. Triage: Manual for traditional fuzzers and built into the scanner for DAST. STAR Fuzzer triages automatically for reachability and impact. Output: Traditional fuzzers return crash inputs and logs. DAST returns a vulnerability report. STAR Fuzzer delivers the file, function and byte-exact payload in a pull request. Fix Validation: Traditional fuzzers need a manual re-run and DAST needs a rescan. With STAR Fuzzer, teams replay the payload. Finding the Bugs LLMs Miss Ask an LLM to audit a parser and it will find the bug classes it already knows from training. The zero-days that survive for years in mature, widely deployed code are the ones that match no known pattern. STAR Fuzzer uses AI to choose targets and build harnesses, but the search that finds the bugs is not an LLM. It doesn’t read the code or guess where a bug might be. It runs the program millions of times, observes what actually happens, and keeps whatever pushes the target into code it hasn’t reached yet. In Back to the Fuzzer II: Finding the Bugs LLMs Miss, Bright Security CTO Bar Hofesh gives four reasons LLM code review does not replace fuzzing for zero-day discovery: Scale: A single fuzzing campaign runs 1 million to 10 million concrete inputs against the compiled target. At about 2,000 tokens per reasoned test case, matching one million executions with an LLM would take about two billion tokens for one target, and each of those test cases would still be a prediction. Speed and Ground Truth: A persistent harness runs hundreds to a few thousand inputs per second per worker, across several parallel workers, so one million executions take minutes to tens of minutes. One LLM reasoning pass over one hypothesis takes several seconds, so one million would take months. The fuzzer also observes real behavior, such as a sanitizer reporting a heap read one byte out of bounds. An LLM has to predict how compiled code behaves, including struct layouts, integer widths, allocator behavior and build flags that aren’t visible in the source. Coverage: STAR Fuzzer’s libxml2 harness reaches about 13,600 code regions in a 5-million-execution run. Coverage feedback shows which regions each input reached, and inputs that reach new code become starting points for deeper search. Code review has no equivalent signal for how much of a large parser it has actually covered. The Unexpected: LLMs, code reviewers and static analyzers look for the same known patterns, so in mature code most of those bugs are already fixed. The fuzzer has no expectations about where a bug should be. The target’s own behavior decides what is interesting. Reproducibility is the other difference. Many plausible-looking LLM findings don’t reproduce, and maintainers need a reproducer before they can act. A fuzzer finding is a reproducer by construction, and Bright Security replays every one in a fresh container before filing it. How STAR Fuzzer Works STAR Fuzzer’s search engine drives any target through a small harness over IPC, so the code under test can be written in C, C++, Rust, Go, Java, Kotlin, Crystal, Node or any language that can read an input and report back. Structure-aware parsing with no per-target grammar: Every seed is parsed into a typed tree of fields by a format-agnostic parsing layer, and the tree serializes back to the original bytes exactly. No grammar or dictionary file is written per target, and each format added to the parsing layer improves every target that uses it. Mutations that get past the first check: Byte-level mutation mostly produces inputs a parser rejects at its first sanity check. Mutating the fields of a valid structured input keeps most executions past framing and validation, so the testing budget is spent on the deeper logic behind them. Evolution with exact replay: Each organism in the population carries a genome, an ordered list of mutations. Each epoch, organisms mutate, run and are scored. Improvements are kept, unproductive changes are reverted, and organisms breed. Replaying a genome reproduces its payload exactly, so every result can be regenerated from a logged random seed. Results from Open-Source Testing Before release, Bright Security ran STAR Fuzzer for two days in an automated pipeline against multiple open source projects, from libraries and applications to low-level binary parsers. 45% of the projects tested had High or Critical issues, and 37% of the vulnerable projects are part of Google’s OSS-Fuzz program. In its first month of availability, STAR Fuzzer produced ten published findings in eight projects written in C, Go, JVM languages and Crystal. Each was replayed in a fresh container before filing: TwelveMonkeys: In the DDS reader, the mipMapCount header field sizes an uncapped array, so a 38-byte file forces a multi-gigabyte allocation (CWE-789). GHSA-fwp7-38wp-mmxw, fixed in 3.15.0. Fider: Upload validation checks an image’s compressed size, then fully decodes it, so a 436 KB PNG expands to about 562 MB. Any authenticated user can trigger it (CWE-409). STAR Fuzzer found it in about 12 minutes, in a release already benchmarked by multiple AI pentesting vendors. GHSA-7cw3-7xh9-529r, fixed in 0.38.0. Crystal standard library (XML): XML::Node#namespace type confusion. ENTITY_DECL is missing from an allowlist, so a field of libxml2’s xmlEntity struct is read as an xmlNs* (CWE-843). GHSA-c267-9cqj-hv9f, fixed in Crystal 1.21.1. Crystal standard library (YAML): YAML::Any#hash overflows the stack on a self-referencing anchor and alias cycle. An 88-byte input causes an uncatchable SIGSEGV (CWE-674). GHSA-fj8w-2r88-5v7w, fixed in Crystal 1.21.1. libsolv: Two vulnerabilities, CVE-2026-104621 and CVE-2026-104622. Details are under embargo. htop: An out-of-bounds read in the /proc/PID/stat parser, caused by a missing bounds check on field 22 (CWE-125). Fixed (#2101). universal-ctags: A null-pointer member access in foreachEntriesInScope (CWE-476). Fixed in commit 780bd7a (#4497). scrimage: An uncapped width-times-height pixel count lets a 35-byte GIF trigger allocation attempts of about 2 GB, and over 8 GB in a second variant (CWE-789). Fixed in 4.6.8 (#893). RIOT OS: An off-by-one in _dirlist_cb (link_format.c). The bounds check runs after the write, so a NUL byte lands one past the buffer (CWE-193). Fix submitted in PR #22879. Several of these bugs sit in code that looks correct on review. The Fider upload has a size limit, but it measures compressed size. The TwelveMonkeys field reads as image metadata, not as a length. The Crystal allowlist looks complete. The YAML hash function is correct for trees, and the bug is a cycle in the input. The full methodology and results are published in two parts: Part I on brightsec.com, and Back to the Fuzzer II: Finding the Bugs LLMs Miss by Bar Hofesh. Where STAR Fuzzer Fits Best STAR Fuzzer by Bright Security is designed for teams whose code handles untrusted input below the HTTP layer, including: Applications that process uploaded files, images, documents or archives Parsers, serialization code and raw binary protocols Non-web software such as embedded and IoT firmware, desktop applications and backend services Codebases with a growing share of AI-generated code that needs testing beyond pattern matching Security teams that want to confirm static analysis (SAST) findings by exercising the flagged functions live It supports C/C++, Rust, Go, Java/Kotlin, .NET, Node, Python and Crystal, and runs inside the customer’s environment with no running application, cloud scanner or authentication required. In persistent mode, harness execution is about 51 times faster than running a new process per input. Executive Perspectives “Fuzzing has always found the bugs that matter most, the ones that take a service down or corrupt memory, but it has been too expensive for most teams to run,” said Bar Hofesh, Co-founder and CTO of Bright Security. “The search was never the bottleneck. The harnesses and the triage were. STAR Fuzzer automates both, so the evolutionary engine we built in 2018 can finally run across a whole codebase.””When Nevada opened the Boulder City Bypass in 2018, it took navigation apps months to start sending drivers down it,” Hofesh added. “LLMs auditing code work the same way. They find the bugs they’ve already seen. The zero-days that have lived for years in mature code are the ones nobody has seen before, and that’s exactly where a fuzzer shines.””Security teams are being asked to keep up with code written faster than ever, much of it by AI,” said Gadi Bashvitz, Co-founder and CEO of Bright Security. “Testing that only looks where bugs are expected won’t keep pace. STAR Fuzzer looks everywhere else, and it hands developers the proof they need to fix what it finds.””AppSec teams keep asking us the same thing: if AI is writing more of our code, what’s testing the parts AI can’t, see?” said Eyal Dror, CMO of Bright Security. “STAR Fuzzer is our answer. In its first month it found ten real bugs in mature open-source projects, and everyone came with the exact input that triggers it. That’s proof a developer can act on the same day, not another alert to argue about.” Availability STAR Fuzzer is available now as a mode of Bright STAR. Learn more and book a demo at brightsec.com/fuzzer. Frequently Asked Questions: What is Agentic Fuzzing? Agentic fuzzing uses AI agents to automate the judgment-heavy steps of fuzzing: selecting which functions to test, writing and verifying harnesses, and triaging results. The fuzzing itself is performed by a search engine that mutates inputs to trigger faults. How is STAR Fuzzer Different from AFL++ or libFuzzer? AFL++ and libFuzzer are fuzzing engines that need engineers to write a harness for each target and to triage crashes by hand. STAR Fuzzer by Bright Security selects targets, generates and verifies harnesses, triages faults for reachability and impact, and delivers findings in a pull request with a reproducing payload. Can LLMs Find Zero-Day Vulnerabilities in Parsers? LLMs are effective at known bug classes, such as an unchecked memcpy length or string-built SQL, but they look where their training says bugs should be. In mature code, the bugs that survive are the ones that match no known pattern, which is where evolutionary, coverage-guided fuzzing is strongest. Why Use a Fuzzer Instead of LLM Code Review? A fuzzing campaign runs millions of real executions against compiled code and observes what happens, while an LLM predicts behavior from the source. Matching one million executions with LLM reasoning would take about two billion tokens per target. Every fuzzer finding also comes with an input that reproduces it. How are STAR Fuzzer Findings Verified? Every finding comes with a byte-exact payload that reproduces it, and Bright Security replays each one in a fresh container before filing it. Any result can also be regenerated from a logged random seed. Does STAR Fuzzer Need Grammars or Dictionary Files? No. A format-agnostic parsing layer turns each seed into a typed tree of fields, so no per-target grammar or dictionary is required. Does STAR Fuzzer Replace DAST? No. DAST tests a running application from the outside in, over HTTP. STAR Fuzzer tests security-critical functions from the inside out, without a running application. Bright STAR runs both, with the same agent and the same pull-request output. What Kinds of Bugs Does STAR Fuzzer Find? Crashes, hangs, out-of-memory conditions and unbounded allocations, memory-safety bugs such as out-of-bounds reads and off-by-one writes, type confusion, stack overflows from recursive input structures, and algorithmic-complexity blow-ups, including decompression bombs triggered by inputs of a few dozen bytes. Which languages does STAR Fuzzer Support? C/C++, Rust, Go, Java/Kotlin, .NET, Node, Python and Crystal. About Bright Security Bright Security is the Agentic Offensive Application Security company, helping organizations secure software at the speed of AI-driven development. Its AI-powered Bright STAR platform integrates security directly into the development lifecycle, automating detection, remediation and validation across web applications, APIs and code, and fitting into both developer and AppSec workflows. Formerly known as NeuraLegion, Bright Security is headquartered in California. Learn more at brightsec.com. Contact CMO Eyal Dror Bright Security Inc eyal.dror@brightsec.com]]> https://www.it-connect.fr/quad9-quitte-la-france-blocage-dns/ 2026-10-26T08:35:00+00:00 2026-10-26T08:35:00+00:00 Florian BURNEL Vous utilisez le DNS 9.9.9.9 ? Le 24 novembre 2026, Quad9 cessera de servir la France plutôt que d'appliquer un blocage judiciaire. Voici ce que l'on sait. Le post Quad9 quitte la France : le DNS 9.9.9.9 sera coupé le 24 novembre 2026 a été publié sur IT-Connect.]]> https://www.helpnetsecurity.com/2026/10/26/witness-open-source-software-supply-chain-tool/ 2026-10-26T08:00:00+00:00 2026-10-26T08:00:00+00:00 Anamarija Pogorelec Witness is an open-source command-line tool that records how software gets built, signs that record, and checks it against a policy before the software ships. TestifySec built Witness and donated it to the in-toto ecosystem within the Cloud Native Computing Foundation, and an open community now maintains it. A compiled binary doesn’t say which machine built it or what code went in. Witness lets a team check which command produced the file, which materials went … More → The post Witness: Open-source framework for verifying how your software was built appeared first on Help Net Security.]]>