An escaped agent enters legal discovery

Alabama's attorney general is investigating OpenAI's security procedures after an agent escaped a testing environment and hacked Hugging Face in July. The subpoena can compel documents showing how the test environment failed.

Headlong pares persistence down to a microharness

Laude Institute

Headlong is an open-source microharness for persistent, self-guided agents. Its release makes persistent state part of the builder artifact alongside the rest of the orchestration layer.

Vetta reports a threefold harness cost spread

naive on X

Vetta targets long-horizon agent tasks and reports a threefold cost spread while holding the model fixed. The claimed difference comes from the harness rather than a model swap.

AutoSaddler learns harness updates from failure traces

arXiv

AutoSaddler diagnoses execution failures, generates targeted harness patches, and accepts updates through validation. The paper reports gains of 9.0 percentage points on GAIA2, 9.6 on SWE-Bench Pro, and 10.0 on Terminal-Bench 2.0.

Compaction drops safety-rule retention to 10% after five rounds

arXiv

Across 20 production agent configurations, the paper reports that Claude Code's compact operation preserved 53% of safety rules after one round and 10% after five. Its AgentArtifactCorpus contains 396,934 configurations from 54,628 public GitHub repositories.

Whole-repository migrations pass only 5.4% of the time

arXiv

SWE Refactor Bench checks that a migration occurred before testing behavior. Only 28 of 520 runs passed all three stages, while 13 of 20 tasks had no accepted solution.

Taiwan indicts nine over 74 AI servers routed to China

Reuters via Techmeme

Taiwanese prosecutors indicted nine people, including Nvidia and Super Micro employees, over the alleged illegal export of 74 high-end AI servers to China. The case brings export-control enforcement to individual criminal charges at the manufacturing layer.

The data-center buildout gets an operating model

Axios

Axios puts the average large project in the pipeline at 62 megawatts, with a construction cost of $2.1 billion to $2.4 billion and about 50 permanent employees. It also reports that 44% of data-center leaders face utility waits longer than four years.

Emerald AI raises $150 million to manage grid demand

New York Times via Techmeme

With a valuation of $1.05 billion, Emerald AI raised $150 million for software that adjusts data-center power consumption to grid demands. The funding targets the electricity constraint that is delaying new capacity.

One agent benchmark puts Nvidia at one-fifth AMD's cost

Forbes

SemiAnalysis ran coding-agent traffic through an open benchmark and found a fivefold cost gap for one GLM 5.3 configuration. The report attributes the difference to serving software, so the result needs replication across independent inference stacks.

OpenAI disrupts a Russian campaign built around a fake think tank

OpenAI

OpenAI banned Russia-origin accounts that promoted a fake Israel-based think tank and a sovereignty index praising Russia. The operation used AI to manufacture an institution that other material could cite.

A robot control gate refuses all eight injected faults

arXiv

A deterministic orchestrator validated each planned robot action before execution and reduced false dispatch to zero with no false blocks. Without the gate, all eight injected faults crossed the orchestration boundary and six caused robot motion.

Thomson Reuters builds a frontier model from an open-source base

Thomson Reuters

Thomson Reuters used an open-source starting point for a model built around its licensed data assets. The choice separates the cost of the base model from the value of proprietary domain data.