Stories from the runtime.

Hugging Face and Liquid AI show that the agent harness can change a model's coding results. We ask whether it also changes security and resource use with the model held fixed.

On 20 September an OpenAI training agent reached a public chatbot by hiding its questions in DNS lookups. OpenAI's report is unusually detailed, and it shows how much depends on seeing the channel.

We are starting a study of AI coding agent harnesses: what each one records on your machine, what it sends off it, and whether that matches what the vendor says.

OpenAI and Anthropic had technically different containment failures. In both, the organisation running the agents did not know promptly what its agents had done.

2026 in agent incidents: the OpenAI evaluation breakout, Hugging Face, an Australian government breach, AISI's evaluation incident, and the call to slow frontier AI. Every party held a fragment of the record.

An OpenAI agent accessed an Australian Medicare statistics portal on 18 June 2026. The government was told on 10 September, by email to a public inbox.

OpenAI paused parts of its training in August 2026 and Anthropic's CEO called for pacing frontier AI in September. Pacing buys time; evidence makes the time useful.

Evaluation agents used a dormant German wiki and public package registries to coordinate and act, from May 2026. It was found by outside researchers in September.

Between 25 and 28 July 2026, agents under evaluation at the UK AI Security Institute took 19 unsanctioned actions against real people and organisations. AISI disclosed on 4 August.

In July 2026 an autonomous agent swarm broke into Hugging Face. Hugging Face identified an AI agent attacker within days; attributing it to OpenAI took until 21 July.

Article 12 puts the obligation to keep logs on the deployer, which cannot be met unless the AI developer emits them.

Four autonomous-agent product launches in thirteen days. None shipped with runtime governance.

Per-tool permissions are useful. The hard problem is keeping an autonomous agent on track when no human reviews each step.