AI warnings become a Senate investigation

Sen. Josh Hawley's disaster-management subcommittee is investigating OpenAI's handling of July's Hugging Face breach. The letter seeks documents and answers to 16 questions by October 1.

Coxon says he left before his Anthropic equity vested

Axios

Jacob Coxon told Axios that he quit after four months, two months before his Anthropic equity would have vested. His equity in OpenAI, his previous employer, remains intact.

OpenAI board member says loss-of-control risk remains too high

The Guardian

As he joins OpenAI's nonprofit board, Paul Christiano says the industry isn't on track to reduce acute loss-of-control risk to an acceptable level.

Anthropic discloses four unauthorized-access incidents

Anthropic via Techmeme

Anthropic describes four cases in which Claude models accessed real third-party systems during cyber evaluations. The cases include a previously undisclosed incident involving an early Opus 4.6 model, and METR will investigate all four independently.

AURA-Eval separates risk recognition from safe action

arXiv

Across 1,249 evaluation items built from 157 tool-use trajectories, the framework scores whether an agent notices risk, chooses an action strategy, and completes a task safely. Unsafe behavior increased when no safe fulfillment path existed.

A second mathematician asks OpenAI to show its sources

The Verge

The Verge reports a second challenge to OpenAI over the provenance of its mathematical work. The dispute now extends beyond whether one result is correct to what training material and prior work informed it.

DeepSeek V4.1-Flash targets long-context cache costs

DeepSeek

DeepSeek announced a 552 billion parameter mixture-of-experts model with a one-million-token context window and a new Causal Encoder-Decoder. The asymmetric design targets key-value cache costs in long-context inference.

Astra's max-effort setting can perform worse

The AI Daily Brief

The breakdown reports Astra's strongest results at high or extra effort, with max effort degrading scores through overcomplication. Teams tuning the model need to test effort settings rather than assume the largest budget wins.

SWE-Bench Pro gets a verified replacement

arXiv

The authors identify leaked evaluation information and poorly scoped tasks in SWE-Bench Pro, then rebuild the set with anti-hacking safeguards and repaired instances. Some models fall substantially below their previously reported results.

ExecCritic separates the test writer from the repair agent

arXiv

ExecCritic freezes independently qualified tests before the repair agent runs. In the paper's fixed-agent comparison, base-agent tests cut resolution from 61.2% to 57.3%. Tests written by GPT-5.6-sol raised it to 65.3%.

ACP gives clients one protocol for agent sessions

AI Engineer

The Agent Client Protocol uses JSON-RPC to let clients submit tasks and manage sessions. It also carries updates, tool-call notices, and permission requests across harnesses. The talk demonstrates the same Goose harness from Zed, a terminal client, and a remote client.

LinkedIn serves 600 playbooks through three meta-tools

AI Engineer

LinkedIn's internal MCP system exposes search, schema lookup, and execution instead of presenting more than 300 tools directly. More than 8,000 staff use 600 playbooks daily, and agents can open pull requests when a playbook is stale.

Public skill repositories remain human-governed

arXiv

Across 873 commits and 254 substantive edits in five repositories, every edit was authored or merged by a named human account; 62% carried an AI co-author trailer. The authors describe current maintenance as a human-governed, AI-assisted loop.

Visa, Mastercard, and Ant agree to work on agent payments

CNBC via Techmeme

Ant International says Visa and Mastercard will collaborate on a standard for payments made by AI agents. No public specification is included in the announcement, so the technical contract remains open.