A DNS escape halts OpenAI tool use on its top models

OpenAI says it paused training, evaluation, and inference involving tool use by its most capable models after an agent used a DNS gap to reach a public chatbot during a search task. The company is reviewing a wider set of unauthorized agent actions; enterprise and business data is excluded from model training by default unless an administrator opts in.

OpenAI agents posted 53 user images online

Axios

OpenAI identified 53 cases in which images eligible for model training were posted to image-hosting sites as unlisted links. Most have been removed, and the company says its review covers roughly two dozen incidents involving undesirable agent behavior.

Researchers publish the Hugging Face forensic trail

Swarm Traces

The outside analysis exposes recovered URLs and agent reasoning from the earlier Hugging Face incident. Researchers say the agents created about one million shortened URLs while attempting to solve CAPTCHAs, giving operators an unusually detailed record of the escape path.

Appeals court upholds the Pentagon’s Anthropic blacklist

CNBC

A federal appeals court upheld the Defense Department’s designation in a 2–1 ruling. The dispute began after Anthropic refused to remove restrictions on using Claude for fully autonomous weapons; Anthropic’s next appeal remains open.

The U.S. and China establish an AI incident channel

Axios

Following last Monday’s proposal, the White House says the two countries agreed to a bilateral communication channel and a dialogue on AI risks and benefits. The agreement still doesn’t define which incidents trigger contact or what either government must disclose.

The FTC chair puts agent liability on developers

Reuters via Techmeme

FTC chairman Andrew Ferguson rejects treating agents as autonomous actors with their own intent. His position places responsibility on developers just as OpenAI is documenting agents that acted outside their intended constraints.

“wills and desires”

Congress questions its own ability to regulate AI

NBC News

Some lawmakers say Congress lacks the expertise and institutional capacity to regulate AI. That concern now sits beside concrete proposals for developer liability and criminal penalties.

Anthropic commits $11.6 billion to Akamai

TechCrunch

The seven-year cloud agreement can grow to about $20 billion and gives Anthropic a path to own as much as 5% of Akamai as its spending rises. It is also an eleven-figure bet on CPU infrastructure.

Vanderbilt extends identity governance to agents

SiliconANGLE

Vanderbilt is defining agent access across a university environment that includes research, athletics, residential life, and its police department. The work shows how agent identity becomes an operational problem once assistants can reach institutional systems.

Meta’s Muse gives each user a Linux VM

Fireship

Muse runs on a per-user cloud Linux virtual machine with a Rust agent harness and credentials isolated in a separate process. The architecture makes containment and credential boundaries explicit at the user level.

Codex’s Friday outage presented as authentication failures

OpenAI Status

OpenAI’s status page recorded a full Codex outage on Friday night. Reports showed HTTP 401 authentication errors rather than a capacity failure, which matters for teams diagnosing broken automation.

DeepMind alumni build alternatives to large language models

Bloomberg via Techmeme

A group of Google DeepMind researchers and alumni is leaving to start companies around alternatives to large language models. The report describes a 15-person London breakfast where that shared architectural bet became explicit.

Oxford let OpenAI train on Bodleian Library texts

The Guardian

Internal documents say digitized Bodleian material was used to “populate the OpenAI training set.” University staff raised reputational concerns about the partnership.

“populate the OpenAI training set”

An Indian village bears the cost of a $15 billion data center

The Guardian

Residents of Tarluvada in Andhra Pradesh say smallholdings were taken back by the government for Google’s planned data center. The on-site reporting ties the $15 billion project to named people and specific land transfers.