▲ OpenAI, Gemini and
Anthropic
US media outlet Axios reported on the 20th (local time) that an artificial intelligence assisting a security team in recovering from a hacking incident refused to cooperate with an investigation into malware.
According to the report, the open-source coding agent “OpenCode” suffered a security breach last weekend, resulting in the leakage of some customers’ application programming interface (API) keys to hackers.
The leaked keys included those for OpenAI, Anthropic, and Google services.
OpenCode stated that the breach was related to a “large-scale distillation attack” and that 1.4% of OpenCode console users were affected.
Dax Raad, co-creator of OpenCode, explained that attackers bombarded the OpenCode system continuously for 48 hours, requiring OpenCode to run its own AI agents around the clock to respond.
OpenAI, Google, and Anthropic explained that they worked with affected customers to revoke exposed API keys and resolve unauthorized usage fee issues.
Writing in a post on social media platform X on the 18th, Raad described the situation, stating that advanced AI models initially helped investigate the incident but then “began to refuse.”
As OpenCode handled code that was potentially malicious during the investigation, safety restrictions embedded in the AI models reportedly blocked the requests, even though the system was trying to defend itself.
OpenCode subsequently gained access to OpenAI’s “Daybreak Blue” program.
This program grants verified cybersecurity defenders access to models with relaxed restrictions.
On the 19th, Raad wrote on X, “We definitely need to be more proactive, and we’ve made really a lot of improvements over the last few days.”
Axios evaluated that the incident highlights the dilemma faced by cybersecurity teams in deciding which and when to use proprietary AI models—which are powerful but come with built-in restrictions—versus open-weight alternatives that offer broader control over deployment and operation.
During the Hugging Face hacking incident, the company discovered suspicious activity on its systems and later linked it to OpenAI’s own agents.
At the time, encountering usage restrictions on cutting-edge models, the security team utilized an open-weight model from China to conduct the investigation.
However, Axios noted that AI companies maintain the stance that access to their most powerful models must be restricted to prevent malicious actors from abusing the cyber capabilities of those models.
Chris Jones, Chief Technology Officer (CTO) at Nightwing, told Axios that cybersecurity teams will need to find a way to balance the need to leverage AI for faster defense with the restrictions designed to keep those very capabilities out of the hands of attackers.
“Over time, we’re all going to have to figure out how to orchestrate this race,” he added.
※ Please note: This article was translated by AI and may contain errors.
Click Here For The Original Source.
——————————————————–
