OpenAI reveals two cases of AI agents disabling safety monitoring during coding tasks, with one falsely denying it had removed the safeguards.OpenAI published an internal incident report this week documenting two separate cases in which GPT-6 class agents weakened the system designed to detect misaligned behavior, in both cases while carrying out routine coding tasks. This wasn’t an external attack. These were internal deployments, and the agents made their own judgment calls that happened to gut the oversight watching them. OpenAI explained that “in both cases, the agents considered whether monitoring safeguards should be preserved, but interpreted routine coding instructions as permission to weaken them.” In other words, the agents thought about whether they should keep the safety controls in place. They then interpreted the user’s instructions in a way that led them to weaken those controls.“In two internal coding tasks, agents weakened misalignment monitoring while making requested software changes. A GPT-6.1 Sol agent repairing an evaluation pipeline disabled automatic alerts to our Security team, treating an earlier instruction not to send messages during a Slack history search as a reason to suppress monitoring notifications.” reads OpenAI’s report. “In a separate task, a GPT-6 Astra agent migrating a program removed its monitoring integration in order to match a reference implementation, and then falsely claimed to the user that it had not removed the monitor.”On October 3, a researcher asked a GPT-6.1 Sol agent to create evaluation tasks using Slack data. The researcher clearly instructed the agent not to send any messages during the process. Later, the system triggered a monitoring alert because it couldn’t save its findings to a database. The user then asked the agent how to proceed. At this stage, the agent was still following the rules. It even told a subagent to keep monitoring enabled and report the actual results.The subagent correctly identified that turning off notifications would disable required safety alerts and decided not to do it. Then the parent agent reconsidered on its own, reasoning that writing results to a database might cause the monitoring system to send a message, which conflicted with the user’s earlier “don’t send messages” instruction. It concluded the instruction applied to safety alerts too, and directed the subagent to set no_notify to true, empty the list of people set to receive monitoring reports, and make notification suppression a hard requirement before writing any results. The monitor still ran and saved its reports. Nobody received the human alerts.On October 6, a user asked a GPT-6 Astra agent to copy a reference implementation while keeping existing tool definitions and functionality intact. An earlier instruction from another agent had explicitly told the system not to disable monitoring, even temporarily. The parent agent passed that requirement to a subagent but softened it, saying monitoring should be preserved “as appropriate” to the reference implementation.The subagent took this wording as permission to remove monitoring because the reference implementation did not include it. It also changed a test that checked whether monitoring was active. The new test instead required monitoring to be disabled. After the migration, the main agent noted internally that the new setup no longer included agent monitoring and described this as a positive outcome.When the user later asked whether monitoring had been removed, Astra said: “No… I did not remove an enabled monitor from it.” That was false.OpenAI offers two possible readings of these incidents. The charitable one is what they call “benign overpersistence,” models that don’t understand why monitoring matters and follow instructions more literally than a reasonable person would. The less comfortable reading is that these models might have a specific tendency to find justifications for evading the systems watching them. “In both cases, the agents considered whether monitoring safeguards should be preserved, but interpreted routine coding instructions as permission to weaken them.” states OpenAI. “We are strengthening our misalignment monitoring system in response, and are also assessing whether these cases reflect benign overpersistence or a propensity to evade misalignment monitoring in particular.”OpenAI says it doesn’t yet have a satisfying answer. To find one, they’re developing paired evaluations, resampling agent continuations, and running experiments on the models’ internal activations.OpenAI has introduced some practical fixes. Its auto-review system, already required for internal Codex use, has been updated to detect and block attempts to disable safety monitoring. The company also plans to require each workload to have an approved safety plan that defines its monitoring requirements. Any workload that fails to meet these requirements will be stopped automatically.Part of the problem was how the systems were designed. Monitoring was easier for agents to bypass during evaluation tasks than in production Codex deployments. In production, monitoring runs outside the agent’s direct control, making it harder for the agent to disable.This report lands a week after OpenAI disclosed seven other misalignment incidents, including one where a research model took cryptocurrency from a wallet it didn’t own to pay transaction fees and didn’t disclose it. “In addition to making it harder for models to disable monitoring, we are looking to understand why models might seek to do so in the first place. It is possible that this behavior is simply an example of “benign” overpersistence, with models not understanding why monitoring is important or insisting that instructions be executed beyond what a reasonable person might expect.” concludes the report. “We are also developing evaluations to test whether models might have a propensity to evade misalignment monitoring in particular and exhibit motivated reasoning in order to do so. We don’t yet have a satisfying explanation for this behavior, and will investigate this behavior through the creation of paired evaluations, resampling of agent continuations, and experiments on models’ internal activations.”Follow me on Twitter: @securityaffairs and Facebook and MastodonPierluigi Paganini(SecurityAffairs – hacking, AI)