OpenAI says prompt injection is unlikely to ever be fully solved

Why the screen matters22 December 2025

Describing its defenses for an AI browser agent, OpenAI wrote that prompt injection, like scams and social engineering on the web, is unlikely to ever be fully solved, and described an automated attacker it built to find new ones.

Position
A vendor statement on AI agents reading web content

What happened

On 22 December 2025 OpenAI published a post on hardening ChatGPT Atlas, its browser agent, against prompt injection. It wrote that prompt injection, much like scams and social engineering on the web, is unlikely to ever be fully solved. It described an automated attacker trained to find injections, including multi-step ones, and an example in which a planted email led the agent to send a resignation letter instead of an out-of-office reply. The findings feed training and monitoring. Earlier that month the UK's NCSC had warned that prompt injection may never be totally mitigated.

The ATG nexus

A vendor at the frontier of AI agents describes the exposure as permanent and managed, which matches how the research treats ATG itself: an endemic condition of the architecture rather than a bug awaiting a patch. As agents act on what they read, the integrity of what reaches them becomes a question of the same kind as the integrity of what reaches a person.

Limits of the comparison

The resignation-letter case was OpenAI's own test, not a real incident.

On this site

Sources

  1. Continuously hardening ChatGPT Atlas against prompt injectionOpenAI, 22 December 2025Primary source
  2. OpenAI says AI browsers may always be vulnerable to prompt injection attacksTechCrunch, 22 December 2025

Back to this event on the timeline