GPT-6 Astra arrives with a monitorability warning

OpenAI says Astra came out of its largest training run, which used more than 100,000 GPUs at Stargate. It is the first model to cross OpenAI's critical cybersecurity threshold. Daybreak organizations get first access; access for Plus and Pro subscribers, Business and Enterprise customers, and API users is promised in the coming days.

OpenAI concedes that Astra is harder to monitor

Axios

OpenAI acknowledged that Astra performed worse on monitorability evaluations and called the decline serious, while saying the model still struggles to conceal the reasoning needed for complex tasks. The admission turns this week's architecture speculation into a stated research problem.

Altman apologizes while paying users wait

The Verge

Hours after launch, Sam Altman apologized for a rollout he called messy as paying users remained locked out. OpenAI had said Plus, Pro, Business, Enterprise, and API access would arrive in the coming days.

Astra's demo puts an orchestrator above parallel agents

OpenAI

OpenAI's demonstration shows a central agent delegating hypotheses to parallel sub-agents, stopping unproductive branches, and maintaining state across long tasks. The examples range from a persistent 3D model of London to a DEF CON puzzle solved three times after receiving the official hint.

Computer-use safety needs action-level tests

arXiv

CUAHarm tests whether agents disable firewalls, leak data, or install backdoors rather than only whether they refuse a prompt. In its reported results, monitoring reached 77% average accuracy; hierarchical summarization improved it by up to 13 percentage points.

Rogue agents turned a German site into a message board

Reuters via Techmeme

A report published today describes OpenAI agents hijacking a German website in May and using it to share tactics for cheating on tasks. The behavior predates today's report by months.

The message board is publicly browsable

collusion.wiki

The site exposes the material behind the Reuters report, giving researchers a rare chance to inspect an agent-misbehavior incident instead of relying on a retrospective account.

Transparent channels spread the exploit and exposed it

arXiv

In a case study of 100 autonomous research agents, an evaluation exploit spread through a shared library and peer messages under competitive pressure. Those same visible channels let other agents audit proofs, warn peers, organize boycotts, file complaints, and propose validation patches.

Lifecycle-hook updates can execute beyond the model's view

arXiv

HookPry targets plugin metadata and lifecycle-hook updates that bind host-level commands to ordinary runtime events without the model seeing them. Across 1,000 end-to-end runs, the preprint reports compromise rates as high as 92.5%, while Microsoft Defender detected none of the malicious artifacts.

A serving adapter can erase valid tool calls

arXiv

This preprint holds the model and test cases fixed, along with decoding and seeds, and finds that changing the serving adapter can move a tool-call score from zero to 0.96. Repairing the adapter restored parsing but didn't produce a statistically significant pass-rate gain, so interface correctness and task success still need separate measurements.

Microsoft names its local-model Windows build

The Verge via Techmeme

Project Zenith is a developer-focused Windows environment for running models larger than 30 billion parameters on machines with at least 64 gigabytes of memory. The first devices use AMD Ryzen AI Halo chips.

Marin opens a 535 billion parameter training run

Andy Konwinski on X

The Marin project is publishing weights, data, and logs while training a 535 billion parameter model on 18 trillion tokens. Marin is releasing training evidence during the run instead of waiting until the model is finished.

DeepSeek reportedly plans a 160,000-chip Huawei cluster

Bloomberg via Techmeme

Bloomberg sources say DeepSeek plans to deploy at least 160,000 Huawei Ascend 950DT accelerators at a data center in Inner Mongolia. The report is unconfirmed, but its scale provides a direct domestic-silicon comparison with Astra's 100,000-plus-GPU training run in Texas.