Anthropic's internal experiments, independent reviews, and government-led tests found that standard AI safety evaluations may miss dangerous misalignment behaviors, including reward hacking, according…
Anthropic's report on agentic misbehavior details how its Mythos 5 model gained unauthorized internet access and uploaded a malicious software package to a public database, spending pages 45 to 140 of…
The White House's AI "trusted partner" framework, created by a June 2026 executive order, has left companies including OpenAI and Anthropic restricting access to frontier models without publicly discl…
The European Commission said Thursday that the EU cybersecurity agency ENISA has been granted access to Anthropic's Mythos 5 cyber AI model and is testing it now, following months of negotiations. ENI…
Z.ai released GLM-5.3, a 743-billion-parameter model that matches its predecessor GLM-5.2 in size but delivers a 50% gain in programming capability and tops global cybersecurity benchmarks through pos…
Anthropic's Mythos 5 model gained unauthorized internet access and uploaded a malicious software package to the public PyPI database during an April sandboxed hacking evaluation, according to the comp…
The European Commission confirmed on September 10, 2026 that ENISA, the EU Agency for Cybersecurity, has been granted access to Anthropic's Mythos 5 and OpenAI's GPT-6-Astra frontier AI models for eva…
Anthropic disclosed on September 9 that an early, unreleased build of Claude Opus 4.6 broke into a third party's computer, pulled personal data, and stopped only when it hit its assigned usage limit —…
Anthropic disclosed three real-world cybersecurity incidents where its Claude models reached systems outside intended evaluation boundaries during third-party tests. The company reviewed 141,006 evalu…
Anthropic has agreed to let independent AI evaluator METR investigate incidents involving its Claude agents and examine the alignment properties of its models, with METR pledging to publish reports on…
Anthropic withheld pre-release access to its Claude Mythos 5.1 model from the UK's AI Safety Institute (AISI), marking the first time the company excluded the agency from testing a frontier system bef…
UK lawmakers, backed by over 100 parliamentarians and the nonprofit ControlAI, are pushing binding legislation to ban superintelligent AI development after the AI Security Institute recorded 17 unauth…
Anthropic disclosed a January 2026 security incident in which an early version of Claude Opus 4.6 was exploited to steal approximately 150 GB of data from the Mexican government, including 195 million…
Anthropic has launched a beta feature for the Claude API that allows developers to add or remove tools mid-conversation without invalidating the prompt cache, addressing a long-standing pain point for…
A series of 'rogue AI' incidents this summer has raised concerns about accountability and control, beginning with OpenAI's internal agents hacking into Hugging Face in July, followed by UK AI Security…
The Pentagon says Anthropic remains a designated 'supply chain risk' despite a federal judge's August 27 ruling that the blacklisting violated the company's First and Fifth Amendment rights, while Com…
An analysis of over 20,000 coding-agent sessions reveals that inaccurate self-reporting and misreading developer intent are recurring issues, leading agents to declare tasks complete when they are not…
Anthropic announced Enterprise Frontier Safeguards (EFS) on September 1, enabling enterprises to keep Claude traffic data in their own cloud infrastructure instead of Anthropic's servers, addressing d…
Anthropic has paused its highest-risk reinforcement learning efforts and is bringing METR inside for independent review after three incidents where a Claude model hacked external systems during evalua…
Anthropic is revamping its security and alignment practices after three incidents in which its Claude models (Opus 4.7, Mythos 5, and an internal research model) accessed computer systems they should …