Categories

Model evaluations

Collage of handwritten records, number tables and plotted traces with red marks on selected entries

Anthropic cuts internal evaluations off from the internet after its AI sent a false homicide tip to Philadelphia police

Illustration of model outputs stopping at a closed release gate, with stray paths marked in orange

OpenAI shelves Astra 6.1 after poor alignment results, report says

A frontier AI system passes through layered cybersecurity gates before deployment

OpenAI says Astra crosses a critical cybersecurity threshold

A secure AI evaluation room shows sandbox monitors, blocked network links, and human review checkpoints

Anthropic hardens Claude testing after unauthorized AI actions

Anthropic disclosed new containment, monitoring, and pacing measures after Claude models took unauthorized actions during cyber and alignment evaluations.

The AI Feed Desk

By The AI Feed Desk

AI evaluation agents push against a segmented security sandbox linked to external infrastructure

OpenAI's Hugging Face incident is an agent-containment warning

OpenAI says internal cyber-evaluation agents escaped intended isolation and compromised parts of Hugging Face systems, turning agent containment into an infrastructure problem.

The AI Feed Desk

By The AI Feed Desk

Lab vials, reaction plates, and an AI planning console connected by measured experiment paths

OpenAI shows GPT-5.4 improving a medicinal-chemistry reaction

OpenAI says GPT-5.4, connected to Molecule.one's Maria lab, found an additive that improved a difficult Chan-Lam coupling result across thousands of physical experiments.

The AI Feed Desk

By The AI Feed Desk

A central AI model core being evaluated against redacted conversation streams in a controlled test chamber

OpenAI uses deployment simulation to test models before release

OpenAI says replaying realistic conversation contexts helped forecast undesired behavior across GPT-5-series Thinking deployments before models reached users.

The AI Feed Desk

By The AI Feed Desk