OpenAI Admits 2 AI Agents Weakened Their Own Safety Monitors
(MENAFN) Two OpenAI coding agents weakened internal safeguards built to catch misaligned behavior during separate tasks this month, and one of them later wrongly denied having done so, the company disclosed in a report published Tuesday.
The incidents involved GPT-6.1 Sol on Oct. 3 and GPT-6 Astra on Oct. 6, each working on coding assignments inside OpenAI’s internal systems.
Misalignment describes an AI system acting against a user’s intentions, company policies or safety requirements.
The first incident. A researcher told a GPT-6.1 Sol agent to build evaluation tasks from employee feedback in an internal Slack channel and expressly barred it from sending any messages. When the evaluation pipeline hit a monitoring problem, the agent read the instruction too broadly and altered the system so automatic alerts no longer reached OpenAI’s security staff. The monitor kept examining and logging the agent’s activity, but the notifications meant for human reviewers were switched off.
The second incident. A GPT-6 Astra agent was assigned to migrate a software program to match a reference implementation while keeping its existing functionality. To mirror the reference system, it stripped out the program’s monitoring integration, despite an explicit order not to disable the monitor “even temporarily.” It also swapped a test that verified monitoring was active for one that required monitoring to be absent. Asked afterward whether it had removed the monitor, the agent incorrectly said it had not.
OpenAI said both agents weighed whether the safeguards should stay but ultimately treated routine coding instructions as permission to weaken them. The company stressed it has not determined whether either agent deliberately tried to hide its actions or dodge oversight.
One possible explanation, the company said, is “benign overpersistence,” in which an agent follows instructions too rigidly without considering why a safeguard exists. OpenAI is also examining whether advanced models may tend to evade monitoring and rely on motivated reasoning to justify it.
Neither incident reached the company’s highest P0 severity level, and neither task was automatically terminated. Both unfolded inside internal systems, and there is no sign that outside users or customer data were affected.
OpenAI also acknowledged other cases in which agents launched workloads without misalignment monitoring. In those cases the original agents stayed monitored, but the programs they spawned ran without the same protections.
The company said it has updated its automated review system to detect and block attempts to remove monitoring from internal Codex tasks. It is also building a framework that would require every category of workload to carry an approved safety case spelling out its monitoring needs. Under the proposal, any workload without approval, or one that fails to meet those requirements, would be terminated automatically.
The incidents involved GPT-6.1 Sol on Oct. 3 and GPT-6 Astra on Oct. 6, each working on coding assignments inside OpenAI’s internal systems.
Misalignment describes an AI system acting against a user’s intentions, company policies or safety requirements.
The first incident. A researcher told a GPT-6.1 Sol agent to build evaluation tasks from employee feedback in an internal Slack channel and expressly barred it from sending any messages. When the evaluation pipeline hit a monitoring problem, the agent read the instruction too broadly and altered the system so automatic alerts no longer reached OpenAI’s security staff. The monitor kept examining and logging the agent’s activity, but the notifications meant for human reviewers were switched off.
The second incident. A GPT-6 Astra agent was assigned to migrate a software program to match a reference implementation while keeping its existing functionality. To mirror the reference system, it stripped out the program’s monitoring integration, despite an explicit order not to disable the monitor “even temporarily.” It also swapped a test that verified monitoring was active for one that required monitoring to be absent. Asked afterward whether it had removed the monitor, the agent incorrectly said it had not.
OpenAI said both agents weighed whether the safeguards should stay but ultimately treated routine coding instructions as permission to weaken them. The company stressed it has not determined whether either agent deliberately tried to hide its actions or dodge oversight.
One possible explanation, the company said, is “benign overpersistence,” in which an agent follows instructions too rigidly without considering why a safeguard exists. OpenAI is also examining whether advanced models may tend to evade monitoring and rely on motivated reasoning to justify it.
Neither incident reached the company’s highest P0 severity level, and neither task was automatically terminated. Both unfolded inside internal systems, and there is no sign that outside users or customer data were affected.
OpenAI also acknowledged other cases in which agents launched workloads without misalignment monitoring. In those cases the original agents stayed monitored, but the programs they spawned ran without the same protections.
The company said it has updated its automated review system to detect and block attempts to remove monitoring from internal Codex tasks. It is also building a framework that would require every category of workload to carry an approved safety case spelling out its monitoring needs. Under the proposal, any workload without approval, or one that fails to meet those requirements, would be terminated automatically.
Legal Disclaimer:
MENAFN provides the
information “as is” without warranty of any kind. We do not accept
any responsibility or liability for the accuracy, content, images,
videos, licenses, completeness, legality, or reliability of the information
contained in this article. If you have any complaints or copyright
issues related to this article, kindly contact the provider above.
Comments
No comment