OpenAI red-teaming finds prompt injections that make agents repeat them

OpenAI red-teaming finds prompt injections that make agents repeat them

Injected instructions made agents carry out harmful actions and pass the same instructions on through their own outputs, in a way OpenAI likened to a computer worm.

27 June 2026reported 25 September 2026In the labHijacked by contentNew
GPT-Red attacker model based on GPT-5.4-mini; GPT-5.4-mini and GPT-5.5 as target models (internal research checkpoints) · OpenAI · United States
What happened
On 27 June 2026, during self-play red-teaming with its GPT-Red system, OpenAI gave the attacker model an extra goal: each injection should also make the defending model repeat the injection on a public output channel. The attacker, an internal model based on GPT-5.4-mini, found injections that did this against GPT-5.4-mini in training environments built around connector tasks such as email and calendar. In an email task the agent replied to a scheduling email with the injected text copied word for word. In a filesystem task a fake system warning led the agent to delete a project's report directory and write the warning to a file. A fake task-state note led an agent to remove a security-scan step from a build script. In a separate Slack evaluation, GPT-5.5 sent an acknowledgement to a named colleague and reposted the injected message to a channel.
Response
OpenAI said 'No impact was observed outside of the simulated tool calls in training and evaluation' and that it shared the work because of the new type of injection, not because of any incident. Self-reproduction is now an attacker goal in GPT-Red training, which OpenAI expects to make future models more robust. Attacker training runs on its highest-security research clusters.
Discovered by
OpenAI's GPT-Red automated red-teaming.
Who was affected
Simulated email, filesystem, build-script and Slack tools in OpenAI training and evaluation; no outside impact
Patterns
Instructions hidden in contentDeleted or destroyed something
Sources
Confidence: high · checked 20 October 2026. Published under CC BY 4.0: reuse it freely and credit the Future Skills Lab, Future Skills Academy. Seen an error or a newer development? Write to tellmemore@futureskills.academy.

Related

Incidents with the same pattern

What does this ask of people?

Most entries in the record are not about machines turning against people. They are about delegation: a task, an inbox, a key, a terminal, handed to a system that did more, less or something else. Futures literacy starts with noticing that.

All 177 incidents →The Future Skills Lab