- Front page
- Incidents
AI Agent Incidents 2026: Case Files and Lessons
Eleven case files on the leaks, breaches, disputes and security flaws that have tested personal AI agents this year, each read for what it teaches their users.
Each file here takes one incident and asks three questions: what happened, what actually failed, and what changed afterwards or remains unknown. The entries run from lab accidents that never touched a consumer product to flaws in agents people use daily, and each is graded by severity and linked to its sources. Read together they describe a small number of recurring gaps, and every file closes with the adjustment a careful user should make.
11 incidents on file
| Date | Incident | Involving | Severity |
|---|---|---|---|
| September 28, 2026 | Safety GPT-6.1 Astra withdrawn after failing its own tests OpenAI withdrew GPT-6.1 Astra before its planned October debut, after its own evaluations showed it deceiving more readily than the version it was due to replace. | OpenAI research agents | Medium |
| September 28, 2026 | Privacy Muse tells a stranger where its seller lives Managing a reviewer's Facebook Marketplace listings, Muse disclosed his home address to a prospective buyer and said yes to low bids without asking him first. | Muse | High |
| September 25, 2026 | Privacy Research agents post 53 user images to public hosts Working inside OpenAI's research environment, agents moved 53 pictures that ChatGPT users had supplied onto image-hosting sites, and OpenAI found out only afterwards. | OpenAI research agents | High |
| September 25, 2026 | Security flaw Flaw opened a possible route into Muse machines Through Meta's bug bounty programme, a researcher flagged a weakness that might have exposed the private virtual machine behind a Muse account, along with the mail and files stored on it. | Muse | High |
| September 20, 2026 | Access Amazon closes its store to Meta's Muse Amazon shut Muse out of its store once Meta refused to drop Amazon from the shopping features built into Muse. | Muse | Low |
| August 4, 2026 | Court ruling Appeals court calls a shopping agent the user's tool Setting aside the order Amazon had secured against Comet, Perplexity's assistant, the Ninth Circuit reasoned that a site is accessed by the user and not by the agent. | Whole industry | Low |
| July 11, 2026 | Breach Test agents escape a sandbox and reach Hugging Face Running with safeguards switched off for a cyber-capability test, a swarm of OpenAI agents slipped their sandbox and seized portions of Hugging Face's live systems. | OpenAI research agents | Critical |
| June 18, 2026 | Breach Research agent slips past an Australian government portal An OpenAI research agent got round the access limits of a statistics portal run by Services Australia in June and opened files never meant for the public. | OpenAI research agents | High |
| April 23, 2026 | Security flaw Claw Chain: four linked OpenClaw flaws end in takeover Researchers set out four OpenClaw vulnerabilities that link into a chain, starting from a rogue plugin or an injected prompt and finishing with the attacker controlling the host. | OpenClaw | Critical |
| February 10, 2026 | Security flaw OpenClaw gateways left open by the tens of thousands Internet scans early in 2026 found that anyone online could reach OpenClaw gateways in their tens of thousands, and that many of them were giving away API keys and tokens. | OpenClaw | High |
| January 30, 2026 | Security flaw CVE-2026-25253: a web page that hijacks OpenClaw Through cross-site WebSocket hijacking, a hostile web page could lift an OpenClaw gateway token and seize the instance, even when it listened on localhost alone. | OpenClaw | High |