Tag archive

AI safety

The “AI safety” archive contains 7 published articles, dated 2026-10-04–2026-11-07. Each article retains its sources and context.

7 stories · AI safety

Filed together

Stories in this thread

AI-generated editorial illustration: a microchip and Petri dish inside nested glass protective enclosures, a disconnected network cable and a magnifying glass represent wet-lab AI evaluation and boundary security, not a real laboratory.
Technology/

OpenAI unveils 100-task BioBasicsBench wet-lab benchmark and partners with Hacktron on cross-service security testing

Built with Ligo Biosciences from raw benchtop data, BioBasicsBench shows GPT-6 Astra reaching 66.7% while still misreading PCR gels; separately, OpenAI expanded Daybreak Red access for Hacktron.

AI-generated editorial illustration of a glowing abstract computing core, layered protective structures and circuit connections
Technology/

Kimi K3.1 Pro: cyber benchmarks do not measure real-world attack rates

Controlled cyber benchmarks do not measure real-world attack success or establish deployment safety. Evaluation conditions, access controls and operational risk need separate scrutiny.

AI-generated editorial illustration: a chip, nested protective enclosures and a disconnected cable represent layered isolation in AI training, not an actual OpenAI facility.
Technology/

OpenAI’s training disclosure puts verifiable isolation at the center of AI safety

Safe frontier training requires more than switching off internet access: permissions, infrastructure and failure handling must preserve a boundary that can be tested.

AI-generated editorial illustration: AI agents and independent monitoring, not an actual scene.
Technology/

OpenAI’s monitoring incidents expose a boundary agents should not control

Two internal coding incidents show why a completed task does not prove that oversight still works: independent permissions, acceptance checks and failure handling matter.

AI-generated illustration of a lock and isolation gate in front of a webpage
Technology/

When AI evaluations reach real websites, responsibility extends to notification and repair

Agent evaluations need permissions that prevent unauthorised actions, followed by disclosure to affected parties and verification that remedies work.

AI-generated editorial illustration of federal AI policy coordination, risk assessment and incident reporting
Technology/

Trump confirms Clayton-led AI task force as federal risk review gets a 120-day timetable

Trump announced the federal AI task force’s leadership on October 4. Reuters, citing the Wall Street Journal, says a 120-day review will assess risks, opportunities, government responsibilities and existing incident-reporting and response capabilities.

Trump hosts an artificial intelligence luncheon in the White House East Room on September 29, 2026
Technology/

White House AI safety pledge leaves the accountability test ahead

September 29 brought voluntary corporate reviews and a separate order changing federal terminology. Official remarks and NIST guidance show what would be needed to assess implementation.

THE DAILY BRIEF

The sharpest read in your feed.

Technology, finance, work and city life, current affairs, energy and industry

SUBSCRIBE Free RSS feed · No email collection