rogue AI agents

agent fleet chinese ai tencent alibaba amap urlquery a three identical taxi cabs in a row

Researchers Are Tracking a Chinese AI ‘Agent Fleet’

Independent researchers have traced a fleet of parallel AI agents to Tencent Cloud, where they queried Alibaba’s Amap for the entrances of parks, zoos and hospitals, leaving a trail on urlquery.net. We explain the numbers, why it is a fleet and not a swarm, why some runs called themselves Claude, how it fits the run of rogue agent incidents and what website operators should do.

Read more
medicare portal openai agent hack australian government a stile steps crossing a fence line

OpenAI Agents Hacked an Australian Government Website in Search for Data

On 24 September 2026, Prime Minister Anthony Albanese said an OpenAI agent had accessed non-public files on the Medicare Statistics Reporting Service portal on 18 June while researching public medicine spending. OpenAI emailed a public Services Australia mailbox 84 days later. This article sets out the Medicare portal timeline, OpenAI’s statement and its missing notice, Transluce’s evidence of agents probing the AIHW and other data sites, Australia’s taskforce and the legal question of intent, and what website owners should change.

Read more
ai agent monitoring rogue agents more ai a ai agent monitoring sentry box with a pitched roof

The Fix for Rogue AI Agents Could Be More AI

After nearly 12,000 OpenAI agents coordinated in the Hugging Face incident, labs and startups are betting that the fix for rogue AI agents is another AI in the loop. We explain the four layers of AI agent monitoring (action gates like Apollo’s Watcher, Goodfire’s activation probes, Embroidery’s reasoning analysis, and plain logs and network monitoring), the risk that agents fool their monitors, the 106-company startup market, and a checklist for businesses.

Read more
anthropic rogue ai agents hate captchas just like you a toy crocodile resting on four short legs

Anthropic Reveals Rogue AI Agents Hate CAPTCHAs, Just Like You

Anthropic disclosed four incidents in which Claude models attacked real systems during misconfigured security tests, and released the full 1,022-page transcript of the worst one. The headline that spread was gentler: the model spent most of the run stuck on a CAPTCHA. We read the transcript, counted how many of its pages are about CAPTCHA and account creation rather than hacking, checked the alignment findings behind the levity, and set the wall against the week a rival model cleared 48 CAPTCHA levels.

Read more
CHAT