Anthropic’s contractor chats bypassed biological safety filters for nearly a year

Anthropic’s biological safety filters were inactive on human-feedback contractor traffic from May 2025 through April 2026, The Decoder reports. The automated checks are designed to block dangerous biological or chemical assistance. The affected workflow involved approximately 50,000 participants and 133 million exchanges.

Anthropic’s risk report says an internal-use flag disabled both blocking and the logging of classifier warnings. The company says it fixed the problem and retrospectively screened the retained conversations. Human reviewers examined 62 flagged transcripts from outside red-team exercises, plus a sample of 30 red-team transcripts, and found no clearly concerning misuse.

Top Stories#

OpenAI describes how developers can embed Codex’s execution system in their apps. Its platform post explains the software surrounding the model: maintaining context, calling tools, streaming progress, enforcing configured boundaries, and requesting approval. The integration choices serve different jobs: codex exec handles noninteractive work, the SDK supports programmatic workflows, and app-server supports persistent conversations and approvals.

The post presents an existing open system as an integration platform. OpenAI also reports that retaining reasoning and compacting context raised GPT-5.6 Sol’s ARC-AGI-3 score from 13.3% to 38.3%, while reducing output tokens sixfold.

OpenAI is testing safety checks across related interactions under Zero Data Retention. Its Private Safety Processing preview addresses misuse that may become apparent only across a sequence of requests. Checking each request individually can miss the pattern, including an agent exceeding its authority.

OpenAI says content remains on infrastructure controlled by the organization using the service. Automated processing returns limited safety signals without exposing prompts and responses to OpenAI personnel. The system is still being tested; the page also retains an exception for images flagged as potential child sexual abuse material.

AutoResearchEval examines where scientific agents fail across a whole research workflow. The first version of the study covers finding evidence, running experiments, analyzing results, writing, and review. It describes 100 tasks across seven domains, with 800 trajectories from eight combinations of agent execution software and models.

The authors identify 45 failure patterns, including recurring failures to check outputs against evidence, revise unsupported conclusions, and reconsider an approach. The study uses an automated judge calibrated against human judgments. It leaves untested whether improved orchestration, the software coordinating the agent’s work, would address those failures.

Top Links#

• CoCounsel Legal (thomsonreuters.com): Thomson Reuters announces general availability of a workflow connecting legal research, analysis, drafting, and verification, grounded in Westlaw and Practical Law.
• AI-assisted cybersecurity reporting (nist.gov): NIST’s draft guide includes structured prompts and simulated organizational files for exploring analysis and reporting against Cybersecurity Framework 2.0 outcomes.
• A smaller matrix multiplication exponent (arxiv.org): An AlphaEvolve-assisted preprint reports a better theoretical bound on how computation grows as square matrices get larger, from 2.371339 to less than 2.371177. This does not establish the speed of practical matrix libraries.