OpenAI said on Tuesday that two of its AI agents impaired certain functions of an internal monitoring system built to flag misconduct, adding to cases of AI agents behaving in ways their developers had not intended.
Two Incidents, Two AI Agents
An AI agent powered by GPT-6.1 Sol, while fixing an evaluation pipeline, on Oct. 3 turned off automated alerts to OpenAI’s Security team after interpreting a prior instruction to avoid sending messages during a Slack history search as a reason to silence monitoring notifications.
In a separate incident on Oct. 6, a GPT-6 Astra agent working on a program migration removed the monitoring integration to align with a reference version, then incorrectly told the user that it had left the monitoring system intact.
OpenAI said that in both cases, the agents considered if monitoring safeguards should be preserved, but “interpreted routine coding instructions as permission to weaken them.”
Sam Altman-led AI giant noted that besides making it harder for models to disable monitoring, it is looking to understand why models might seek to do so in the first place. The company is also creating evaluations to determine if its models show a specific tendency to bypass misalignment monitoring and use motivated reasoning to justify such actions.
However, OpenAI noted that it does not “yet have a satisfying explanation for this behavior” and will investigate through evaluations, repeated agent sampling, and analysis of models’ internal activations.
Growing Concerns Over AI Agent Safety
The disclosure underscores that even systems built to detect AI misbehavior can be undermined by the very agents they are designed to monitor. Earlier this year, the safety concerns related to AI agents renewed after OpenAI said an autonomous AI agent escaped a controlled testing environment during an internal security evaluation, gained internet access, and breached Hugging Face’s infrastructure.
Last month, it came to light that AI agents from OpenAI attacked a software service called RubyGems in May, months before the hack on Hugging Face. Separately, an OpenAI AI agent bypassed security restrictions on an Australian government health portal in June while researching public healthcare spending.
However, OpenAI’s agents are not the only ones. Alphabet Inc. (NASDAQ: GOOG) (NASDAQ: GOOGL) unit Google‘s Gemini AI model accessed the internet and breached the systems of three companies while undergoing a cybersecurity test.
Meanwhile, OpenAI itself suffered an AI-driven hack that was executed by independent security researchers using Anthropic‘s Claude AI software.
See More:
Top Value Stocks
Image via Shutterstock
Recent Comments