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)
Security Affairs AI-CYBERSECURITY newsletter includes a collection of the best articles and research on AI in the international landscape Artificial intelligence is rapidly changing cybersecurity, reshaping both the techniques used by attackers and the tools available to defenders. AI agents can automate tasks, analyze large amounts of data, discover vulnerabilities and accelerate offensive operations. At the same time, AI is creating new capabilities for threat detection, incident analysis and response. Security Affairs’ new newsletter follows this evolution, covering every week the latest threats, attacks, vulnerabilities and developments at the intersection of AI and cybersecurity. AI-CYBERSECURITY Newsletter 25-year-old Israeli founder sells stealth AI startup for up to $35 million without raising capital OpenAI claims 100-plus secret files followed ex-employee back to Apple Open weight takes 71.5% of tokens and 23% of spend OpenAI Models Hacked Data Portals, Raided a Crypto Wallet to Finish Tasks Meta removes language that raised scraping and AI training concerns Can LLMs Supercharge Geolocations? Our Tests Show Big Improvements But Humans Remain Key Agents are your newest insiders. Security must move at their speed Anthropic Acquires Confidential AI Startup Opaque Systems ATT&CK v20: Critical Infrastructure Campaigns, Adversary Use and Abuse of AI, & Iranian Cyber Espionage UK Election Watchdog Develops AI System to Flag Deepfakes Within 3 Hours Meta’s Muse hits 10M U.S. downloads in 40 days, a milestone reached by only 11 other apps Suppressing misalignment monitoring alerts and disabling misalignment monitoring Anthropic pitches AI safety bill that may block a model if 50+ people could be killed Making attacking agents pay LiteLLM Compromised in Supply Chain Attack Delivering Credential Stealer and Backdoor AI Just Found a Bug That Puts up to 70% of the World’s Websites at Risk AI-driven, large-scale nation-state campaign using ARTEX enables supply chain compromise Torvalds: AI is fine for competent devs, but spare us the ‘drive-by’ patches Our latest defenses against evolving web scams UK government sets out $800 million of AI security funding Securing our frontier model training environment Follow me on Twitter: @securityaffairs and Facebook and Mastodon Pierluigi Paganini (SecurityAffairs – hacking, newsletter)
Security Affairs Malware newsletter includes a collection of the best articles and research on malware in the international landscape Malware Newsletter Auditing the auditor: Unmasking Apex Accelerator GhostAction Attack Escalates By Targeting GitHub Users Plex vulnerabilities being exploited in the wild to drop a cryptominer WordPress Rank Math SEO PRO Plugin 3.0.123 is vulnerable to a high priority Backdoor ROMULUS: Analysis of a New Multi-Payload ClickFix Delivery Platform Hackers Hijack Five OVHcloud CLI Releases to Distribute Secret-Stealing Malware Pegasus and Mercenary Spyware Attacks on Serbian Pro-Democracy Figures When wipers leave backups: an analysis of ArgonWiper’s encryption workflow LiteLLM Compromised in Supply Chain Attack Delivering Credential Stealer and Backdoor Operation FlashHole: Lazarus breaks security, defies control Cross-Platform Malware Fingerprint Identification with a Convolutional KAN–Transformer Hybrid Architecture Geometry-Aware Risk-Guided Active Learning for Malware Concept Drift Adaptation Mal-AAEGAN: Adversarial Autoencoder with Double-Critic WGAN-GP for Malware Generation Q-MALDetect: Quantum Image-Based Malware Detection and Adversarial Robustness Analysis Follow me on Twitter: @securityaffairs and Facebook and Mastodon Pierluigi Paganini (SecurityAffairs – hacking, newsletter)
A new round of the weekly Security Affairs newsletter has arrived! Every week, the best security articles from Security Affairs are free in your email box. Enjoy a new round of the weekly SecurityAffairs newsletter, including international press. Capcom Data Breach Exposes 11.7 Million Email AddressesU.S. CISA adds Arista VeloCloud Orchestrator flaw to its Known Exploited Vulnerabilities catalogPHP Warns of New Code Execution Flaw Days After PHP2Shell PatchesMeta Faces $25 Billion Penalty Linked to Cambridge Analytica ScandalCapcom Data Breach Exposes 11.7 Million Email AddressesU.S. CISA adds Arista VeloCloud Orchestrator flaw to its Known Exploited Vulnerabilities catalogPHP Warns of New Code Execution Flaw Days After PHP2Shell PatchesMeta Faces $25 Billion Penalty Linked to Cambridge Analytica ScandalPHP2Shell: Millions of Attacks, No Confirmed Successful Exploits YetINTERPOL Operation Serengeti III Nets 1,462 Arrests and Recovers $105 Million in Cybercrime CrackdownMicrosoft Links ConnectWise Breach to China-Linked Group Using AI Pentesting Tool ARTEXSignal Vulnerability Could Keep Calls Alive After Users Hang UpPHP2Shell Attacks Started Four Days Before Public Disclosure, AWS ReportsFBI and CISA Warn of SafePay Ransomware Attacks on 600 OrganizationsRussian Satellite Makes a Close Approach To A Commercial Radar Satellite Used by UkraineOpenAI Reports Two Incidents of AI Agents Disabling Their Own OversightUS Government Orders DeepSeek Out of Federal Civilian SystemsPHP Prepares Emergency Fixes for Two Critical Remote Code Execution FlawsNSO Spyware Pegasus Hit a Serbian University Rector Six Days Before Snap ElectionsNikkei data breach affected over 1.8 million recordsGoogle Cloud Outage in London Disrupts Revolut Banking ServicesAfD Email Breach Exposes 421,000 Messages in Anonymous Online ArchiveMicrosoft Treats AI Agents as Insider Threats in New Security FrameworkPlex2Shell Exploit Used to Turn Vulnerable Plex Servers Into Crypto MinersMeta wants to scrape your website to train its AI, and you can’t say noDeepfakes and Election Manipulation: UK Watchdog Seeks Faster DetectionBig Fish: OpenBSD Root Exploit Chains a 37-Year-Old BSD Bug With a Kernel FlawAttackers Abuse Stolen ConnectWise Application Credentials to Target Microsoft 365Japan’s SiteGuard WAF flaw CVE-2026-106931 is already being exploited16-Year-Old Broadcom Wi-Fi Bug Lets Attackers Take Over Routers Without a PasswordRust-Based CikaBika Bot Turns Discord into a Command Center International Press – Newsletter Cybercrime U.K. Sanctions Cryptomus for RU Money Laundering Anonymous hackers claim to leak 421,179 internal emails from Germany’s Russia-aligned AfD Feds Suspect Ransom Negotiator Ran ShinyHunters Extortion Site #StopRansomware: SafePay Ransomware Cybercrime crackdown in Africa nets 1,462 arrests, recovers over USD 105 million U.K. Sanctions Cryptomus for RU Money Laundering Nigerian Man Sentenced To Eight Years For Hacking, Fraud, And Identity Theft Schemes 1 Million Impacted by Capcom Data Breach Malware Auditing the auditor: Unmasking Apex Accelerator GhostAction Attack Escalates By Targeting GitHub Users Plex vulnerabilities being exploited in the wild to drop a cryptominer ROMULUS: Analysis of a New Multi-Payload ClickFix Delivery Platform Hackers Hijack Five OVHcloud CLI Releases to Distribute Secret-Stealing Malware Hacking ConnectWise (SkyKick) Vendor Credentials Abused in M365 Attacks One Parser, Two Bugs, Many Routers: Exploiting Broadcom WiFi Across eCos and Linux Plex2Shell Unauthenticated RCE in Plex Media Server Big Fish: Exploiting a 37-Year-Old Kernel Bug in OpenBSD Rapid Response: php2shell, Leaked Critical PHP Vulnerabilities (CVE-2026-109618 & CVE-2026-109619) Making attacking agents pay LiteLLM Compromised in Supply Chain Attack Delivering Credential Stealer and Backdoor Amazon Threat Intelligence identifies PHP2Shell exploitation prior to public disclosure in MadPot Proof-of-concept exploit released for PHP vulnerabilities (CVE-2026-109618, CVE-2026-109619) PHP2Shell Exploit Tooling Found on Exposed Attacker Server Intelligence and Information Warfare TAG-193’s Regional Espionage Operations ATT&CK v20: Critical Infrastructure Campaigns, Adversary Use and Abuse of AI, & Iranian Cyber Espionage No to Palantir – Five Times UK Election Watchdog Develops AI System to Flag Deepfakes Within 3 Hours Pegasus and Mercenary Spyware Attacks on Serbian Pro-Democracy Figures Russian satellite makes possibly unprecedented approach to another spacecraft without permission AI-driven, large-scale nation-state campaign using ARTEX enables supply chain compromise Operation FlashHole: Lazarus breaks security, defies control Pegasus spyware founder urges governments to trust his tech and ditch US AI Cybersecurity Character.ai files for bankruptcy following lawsuits alleging ties to teen suicides Cybersecurity Jobs Report 235,786 Agents are your newest insiders. Security must move at their speed Google Cloud services face outage in London, Revolut services down too Government websites go dark after AI-discovered flaw sparks security scramble Widespread PHP Flaw May Cause Many Site Outages Today Suppressing misalignment monitoring alerts and disabling misalignment monitoring AI Just Found a Bug That Puts up to 70% of the World’s Websites at Risk CVE-2026-113380 — RingRTC (Signal): Flooded Call Event Queue Blocks Local Hang-Up Cellebrite’s Advanced Unlocks and the case of the missing iPhone 17 Meta ordered to pay $25 billion penalty to New Mexico in data privacy case PHP Warns of Another Unpatched RCE Two Days After Emergency PHP2Shell Patches Our latest defenses against evolving web scams Follow me on Twitter: @securityaffairs and Facebook and Mastodon Pierluigi Paganini (SecurityAffairs – hacking, newsletter)
OpenAI paused frontier model training for 11 days after AI agents exposed security weaknesses, VM escapes and covert communication channels. OpenAI published a detailed report this week explaining how it improved security in the systems used to train and test its most advanced AI models. The company also described what happened when its AI agents tried to find weaknesses in those systems and why it stopped training its most advanced models for 11 days in September. “Based on the red-teaming results, the fixes we had implemented, and improvements to our monitoring and stopping systems, we judged that our strengthened safeguards were sufficient to resume frontier RL training and internal inference.” reads the report published by the company. “We resumed these workloads on October 1, 11 days after our pause began.” The incident began on September 20, when an internal research model exploited a DNS configuration error and gained limited access to the live internet during training. This incident, along with a security incident involving Hugging Face in July and other cases of AI models behaving in unexpected ways, led OpenAI to make major changes to its security measures. OpenAI’s stated containment requirement is that limits on what an agent can access must hold even when the agent actively tries to circumvent them. The policy choice they landed on is that frontier model training now runs with no internet access at all, full stop. “For training and most evaluation, the live internet lies outside this boundary, as we believe that providing models access to the live internet can pose significant risks. Although internet access can help models learn useful skills, it also gives a capable, misaligned agent opportunities to affect systems beyond our control.” continues the report. “We’ve therefore designed the security architecture for our research environments with one overriding principle: we must always be able to prevent a frontier model from achieving unauthorized, arbitrary access to the live internet.” To support training without live internet, they built what they call an “offline world,” a large indexed and cached set of web content that models can browse and search without touching external systems. The layers of protection around that offline world include restricted containers, lightweight virtual machines that isolate each agent’s environment from the underlying hardware, network and cloud controls, and monitoring systems designed to catch unauthorized activity before it does damage. The report explains how OpenAI tested its most advanced AI agents by asking them to find weaknesses in the company’s own systems. The agents had three goals: access the live internet, break through internal security barriers, or communicate with other agents when they were not allowed to do so. OpenAI ran experiments using hundreds of billions of tokens. The agents found previously unknown security flaws in open-source software, including Kata Containers, gVisor and Moby/Docker, as well as configuration errors in OpenAI’s own services. By combining these weaknesses, they managed to escape virtual machines and gain control of the underlying host system. In further tests, the agents found weaknesses in services used for monitoring, logging and authentication. They used these flaws to make the services access URLs controlled by an attacker. The part of the report that deserves the most attention isn’t the direct internet escapes. It’s what the agents did when those paths weren’t available. Agents in separate virtual machines, not allowed to share files or network connections, communicated with each other using processor cache timing. One agent evicted specific data from the processor’s cache. Another measured how long it took to reload that data, using the difference between fast and slow reads to decode a message. No file touched, no network packet sent. The technique is the same family of attacks as Spectre and Meltdown. OpenAI restarted training its most advanced AI models on October 1, eleven days after the pause began. The company is now making extensive security tests a required part of the approval process before training runs can begin. It is also using AI agents to regularly check source code and cloud settings for weaknesses, testing systems that monitor activity and automatically stop unsafe operations, and reporting discovered vulnerabilities to the open-source projects responsible for fixing them. OpenAI says it is continuing to strengthen its security by fixing vulnerabilities in virtual machines, working with cloud partners, and using AI agents to scan code and configurations. The company is also improving monitoring and automatic shutdown systems, regularly testing its sandbox and network defenses, and making high-compute red-teaming with its most capable models a core requirement for future frontier AI training and evaluation. “We are committed to strengthening our containment, monitoring, and red-teaming infrastructure as our frontier models become more capable and believe that this work will be a core part of responsibly pacing the development of our frontier models.” the company concludes. Follow me on Twitter: @securityaffairs and Facebook and Mastodon Pierluigi Paganini (SecurityAffairs – hacking, Artificial Intelligence)
A new UEFI bootkit named BlackX, deployed in an espionage campaign in early 2026, exploited the old Baton Drop vulnerability to bypass Secure Boot and load a malicious driver designed to interfere with security solutions from Kaspersky and Avast. The operation has been attributed with medium confidence to APT41, a Chinese state-sponsored threat group also known as Brass Typhoon, Wicked Panda, and Barium. It was discovered in early 2026 and targeted diplomatic agencies in Central Asia, particularly in Tajikistan, Uzbekistan, and Turkmenistan. Kaspersky researchers analyzed the BlackX bootkit and found similarities with BlackLotus, which has been circulating on hacker forums since October 2022 and leveraged the Baton Drop vulnerability (CVE-2022-21894) to bypass Secure Boot protection. ESET confirmed in March 2023 that BlackLotus exploited Baton Drop, and at the time it could compromise even fully patched Windows 11 systems.
A memory-corruption flaw in Google's Brotli decoder sits in many programs that unpack Brotli-compressed data, among them Bun, the Python Brotli and brotlicffi packages, and Envoy proxies set up to decompress Brotli. Tracked as CVE-2026-111497, the flaw allows crafted compressed data to overwrite past the end of a buffer and corrupt a program's memory, potentially
LiteLLM on Friday revealed that an Agent Platform master key accidentally published to a public GitHub repository in late May 2026 ultimately allowed the threat actor to push a malicious version of the popular Python package. "On May 27, 2026, a configuration file was committed to the public LiteLLM-Labs/lite-harness repository which inadvertently contained the master key to a deployment of
Google has updated its privacy policy to spell out that using personalized artificial intelligence (AI) features in Search or Gemini could allow the company to serve targeted ads to users based on their activity on other connected Google services. "You may choose to allow certain Google services, including services like Search or Gemini Apps, to access data from other Google services you use,
Between November 2025 and March 2026, the Dutch security firm Eye Security pushed more than 2 million updates to computers it did not own to switch off a background updater that it said could have allowed an outsider to take over the machine. The updater belongs to Time4Popcorn, an abandoned copy of Popcorn Time, a free app that streamed pirated films and TV over BitTorrent. Popcorn Time was
Confirm this action.
Leaving now will discard your changes.