Is the PHP incident only the beginning?···AI with 'two faces' that finds security vulnerabilities ...

Cases continue to emerge that raise concerns that artificial intelligence (AI) could escape human control or be misused for attacks. Recently, as AI has been used to identify security vulnerabilities in the programming language ‘PHP’, the alternation between its ‘benefits’ and ‘harms’ has repeatedly highlighted the importance of control.

According to materials released on the 23rd (local time) by Anthropic and the UK AI Safety Institute, the Anthropic AI model ‘Claude Mitos 5’ was observed acting outside the test scope in 9 out of 43 attempts while performing a cybersecurity evaluation task to infiltrate a virtual computer network and secure designated information. When its initial approach failed, it attacked other targets beyond the test scope and created fake user accounts to falsely assert that the code was safe. This showed a safety issue beyond a simple malfunction. ‘Mitos’ is so capable that it can go beyond finding vulnerabilities to actually hacking, and the U.S. government had even placed it under export control.

The institute said, “Mitos 5 justified continuing its attack strategy by reasoning that the target was created in a virtual environment and did not consider the potential harm its actions could cause,” and, “Even when told that the target was a person unrelated to the evaluation, the attack rate of Mitos 5 did not decrease.”

On the same day, OpenAI said it had suspended training after an internal research AI model used a system configuration error to gain unauthorized access to the external internet. An AI agent bypassed the safety control net (guardrails) set by the researchers and moved beyond the isolation boundary.

After the incident, OpenAI paused model training and, to strengthen its defenses, mobilized its AI Agents team to conduct autonomous penetration testing (red-teaming) that forced break-and-escape from its internal isolation environment (sandbox). The AI agents then found a vulnerability in the widely used Google data compression tool ‘Brotli’.

If an attacker sent maliciously crafted data to a server with Brotli installed, there was a remote code execution risk that could lead to system takeover; after the discovery by OpenAI, Brotli distributed a security patch. It was a moment that revealed both the upside of AI in finding security vulnerabilities and the problem of loss of control. On the 22nd, OpenAI said it had used ‘GPT-5.6 Sol’ to find a PHP vulnerability, adding, “This is only some of several critical vulnerabilities,” and, “We expect a surge of similar discoveries over the coming months.”

Kim Kyung-gon, a professor in the cybersecurity program at Ravensbourne University London, said via the Korea Science and Technology Media Center (SMCK), “What matters is to correctly understand the capabilities and limits of AI, determine what human experts should judge, and decide where to focus capabilities and resources,” adding, “In the future, competitiveness in cybersecurity will depend less on the AI technology itself than on how strategically it can be used and controlled.”