Daily AI news, sourcedOfficial news, first rumor reports and substantive progress. Each entry keeps the information published at that time.
- OpenAI plans its next internal security sprint with Cathedral Systems
OpenAI said it will run another internal security sprint in which Cathedral Systems researchers use OpenAI models alongside OpenAI teams to test defenses. OpenAI also said it will raise bug-bounty payouts for vulnerability discovery by more than 4×; the sprint has not started yet.
- OpenAI publishes an updated Model Spec
OpenAI published an update to its Model Spec, the document that guides model behavior, organized around five core values: faithfulness to human intent, harm minimization, honesty and transparency, human agency, and care. The update adds a faithfulness chapter that frames model behavior as a principal-agent problem, revisits chain of command and social norms defaults, and retires most prescriptive style guidance in favor of letting style adapt to user preferences. OpenAI also renamed the assistant to "GPT" and said it is exploring training methods that teach the spec's core values directly.
- OpenAI details its layered monitoring for frontier model misalignment
OpenAI described the layered monitoring system it uses to catch frontier-model misalignment and unauthorized agent behavior, spanning training, evaluation, internal use, and external deployment. The company said replaying September's Hugging Face incident under the strengthened monitoring would have raised alerts about 30 minutes after the agent escaped its sandbox, more than 30 hours before HF's own systems were compromised. OpenAI also said independent red teams from METR and Redwood Research have been stress-testing the system since July.
- OpenAI describes stronger containment for its frontier model training environment
OpenAI outlines how it hardened containment for frontier model training and evaluation, including offline environments and high-compute red-teaming, to keep training workloads more tightly sandboxed.
- OpenAI says METR and Redwood Research piloted an embedded eval of its internal monitoring
OpenAI says it invited METR and Redwood Research this summer to pilot an embedded evaluation of its internal monitoring system, giving outside researchers deeper access than typical external evals.
- OpenAI and Poland’s NASK partner on AI safety and security research
OpenAI and Poland’s NASK are partnering on AI safety, security and technical research to strengthen safeguards and support a more resilient digital ecosystem.
- OpenAI explains how it holds distribution partners to consistent safeguards under its Usage Policies
OpenAI describes how it works with distribution partners to keep safeguards consistent, support responsible access to its models and clarify what its Usage Policies expect.
- OpenAI publishes part two of its democratic governance of frontier AI series, a blueprint for empowering cyber defenders
OpenAI outlines a blueprint for empowering cyber defenders with frontier AI, alongside stronger safeguards, modernized cybersecurity and more resilient public institutions.
- OpenAI says it disrupted two AI-enabled false-front influence operations using fake journalists and a think tank
OpenAI reports disrupting two AI-enabled influence operations that used false-front journalists and a think tank to spread geopolitical messaging.
- OpenAI lays out safety-case approach for frontier AI training
OpenAI published 'Towards safety cases for frontier AI training', outlining how it thinks about securing frontier RL training runs, per its official X account.
- OpenAI details textGrain watermarking for EU AI Act, detector access starts with researchers
OpenAI published its approach to text watermarking under the EU AI Act, built on textGrain — an invisible statistical signal in the model's word choices. Starting immediately, API customers globally can opt in for select models (off by default), eligible ChatGPT and Codex output in the EU gets the watermark over the coming weeks, and detector access opens to approved researchers and expert organizations only. At a 1% target false-positive rate, detection worked on about 80% of 200-token passages and 95% of 400-token passages, but replacing 25% of words cut detection to 17%. OpenAI says textGrain matched or exceeded SynthID for text in its tests and plans to open-source the technology.