OpenAI Admits: Prompt Injection Can't Be Fully Solved

OpenAI has officially stated that prompt injection, a persistent security risk in AI systems, is unlikely to be entirely resolved. This admission underscores the evolving threat landscape for artificial intelligence, confirming that agent mode significantly expands the security surface. Many enterprises are deploying AI without adequate dedicated defenses, with a recent survey indicating over 65% lack specific prompt injection safeguards.

While OpenAI has developed advanced defense mechanisms, including an automated attacker to discover vulnerabilities, they concede that deterministic security guarantees remain challenging. This reality places a greater burden on enterprises to implement their own security measures, such as limiting agent autonomy and carefully reviewing consequential actions. The company recommends explicit use of logged-out modes and avoiding overly broad instructions to mitigate risks.