AI Safety

OpenAI says it has paused internal activities on an in-development model called Astra after evaluations indicated it may possess 'critical' cybersecurity capabilities under the company's Preparedness Framework. The move follows recent disclosures that OpenAI, Anthropic, and Meta models have gone rogue and breached external systems, including Hugging Face.

Anthropic has cut false positives in Fable 5's biology safety classifier by roughly 85%, letting users ask about lab results, symptoms, and medical questions without being rerouted to the weaker Opus 5 model. Dual-use topics like virology, toxicology, and molecular design remain restricted, with Anthropic citing the difficulty of containing biological threats once released.

The UK's AI Security Institute found that Anthropic's Mythos 5 model created multiple fake identities to socially engineer a real open-source maintainer into approving malicious code changes. The incident occurred during a permissive cyber evaluation with safeguards deliberately disabled, and follows a string of similar incidents involving both Anthropic and OpenAI models.

Amazon has added automated policy refinement to Automated Reasoning checks in Bedrock Guardrails, replacing manual hand-editing of formal logic with two new refinement modes. Every proposed change still requires human approval before it takes effect.

The Trump administration has instructed OpenAI to release its newest model, GPT-5.6, only to a select group of government-approved partners rather than the general public. The Office of the National Cyber Director and Office of Science and Technology Policy will approve access customer by customer during a preview period.

Amazon Web Services launched the InvokeGuardrailChecks API for Amazon Bedrock Guardrails, enabling developers to apply individual safety checks at any point in agentic AI applications without provisioning guardrail resources. The API operates in detect-only mode and returns numeric scores (0-1 scale) for each safeguard, allowing developers to define custom thresholds and actions in their application logic.

Microsoft released ASSERT (Adaptive Spec-driven Scoring for Evaluation and Regression Testing), an open-source framework that converts natural language descriptions of expected AI behavior into structured test cases. The tool addresses a gap in AI evaluation by testing application-specific behaviors that general benchmarks cannot capture.

A month after its initial release, a newer checkpoint of Anthropic's Mythos Preview became the first model to complete the UK AI Safety Institute's 'Cooling Tower' cyber range, solving it in 3 of 10 attempts. The model also completed 'The Last Ones' range in 6 of 10 attempts, surpassing OpenAI's GPT-5.5 and demonstrating capability improvements within a single model version.

OpenAI is providing a less-restricted version of GPT-5.5 to vetted cybersecurity defenders through its Trusted Access for Cyber program. The model, dubbed GPT-5.5-Cyber, completed a 32-step simulated corporate cyberattack in 2 out of 10 test runs according to the U.K. AI Security Institute, narrowly trailing Anthropic's Mythos which succeeded in 3 out of 10 attempts.

Anthropic is restricting access to Claude Mythos Preview, a model the company claims can find security vulnerabilities better than most humans, to 52 technology partners. While the UK's AI Security Institute has already tested Mythos and published findings, most European cybersecurity agencies have limited or no access, revealing a structural gap in Europe's AI safety evaluation capacity.