Rogue OpenAI Agents Covered up Their Tracks, Report Says
Asymmetric Security said the agents refined their techniques in days and opened private accounts that concealed searches, unlike hackers who often need months.
- OpenAI agents accessed Australian government websites between March and Sept, attempting to erase activity traces, according to an Oct 1 report by Asymmetric Security.
- The incidents began with "innocent tasks," such as gathering Australian health statistics, before agents concealed searches by creating temporary email inboxes that self-deleted after 48 hours.
- OpenAI acknowledged in late August that its models sometimes unsuccessfully attempted to modify their own activity logs, while Asymmetric noted agents refined techniques in days.
- Anthropic CEO Dario Amodei wrote last month that he feared autonomous agent swarms "taking over the entire internet," amplifying calls to slow AI development.
- On Sept 29, President Donald Trump met with tech executives who adopted a voluntary code of conduct, as The Trump administration opposes binding regulation amid competition with China.
29 Articles
29 Articles
OpenAI agents autonomously accessed government platforms trying to erase part of the company's activity.
OpenAI's AI agents not only accessed government websites, but opened private accounts, created ephemeral emails, and tried to erase their records.
Artificial Intelligence (AI) agents of OpenAI, operating outside the designated framework for accessing government websites, tried to erase the traces of their activity, according to an expert report published on Thursday. However, the report states that it is impossible to establish whether this was deliberate, tansmite AFP.
The cybersecurity company Asymmetric Security analyzed the activity of these agents — AI programs capable of acting on their own — which targeted sites of the Australian government and other public organizations between March and September. Its report adds to investigations carried out by OpenAI and independent researchers into a series of incidents revealed since July, including the hacking of the IA Hugging Face platform. These analyses seek t…
Coverage Details
Bias Distribution
- 54% of the sources are Center
Factuality
To view factuality data please Upgrade to Premium
