As news spread that Anthropic is testing a new security model known as Claude Mythos, shares of security companies again fell. The move reflected concerns that AI could reduce demand for products from existing security firms.

Foreign media reports said Claude Mythos scored much higher than the current top model, Claude Opus 4.6, across several assessments including software coding, academic reasoning and cybersecurity.

Ed Sim (에드 심), founder of Boldstart Ventures, made clear his view that advances in AI models are an opportunity for the security industry, not a threat. He said cyber threats are growing alongside the evolution of AI models. He also highlighted that Anthropic itself assessed Claude Mythos as creating unprecedented cybersecurity risks. He said the warning was not that the model itself is dangerous, but that if it is abused, the level of cyberattacks could rise to unprecedented levels.

Sim said, "Claude is not killing the cybersecurity industry, but rather growing it exponentially," adding, "AI-based attack techniques are already happening in the real world."

He cited a recent case in which LiteLLM, a standard adapter library used to call major AI services such as OpenAI, Anthropic and Google, was exposed to a supply-chain attack.

He said the tool used by the attacking team TeamPCP, called "hackerbot-claw", automated the selection of attack targets by using AI agents. He said researchers see it as one of the first cases in which an AI agent was actually deployed in a supply-chain attack.

Sim also stressed that the attack was discovered by a person, not AI.

He said FutureSearch developer Callum McMahon (캘럼 맥마흔) learned of the breach after seeing system crashes caused by a malicious payload. He said automated scanners failed to detect it after being fooled by legitimate admin credentials and a valid pip hash.

Sim said the attack is another example showing the importance of a layered approach to security. He said large language models alone have limits in defense. He said, "Even if a foundation model finds 500 vulnerabilities in a codebase, whether those are real vulnerabilities, whether they have been reported, and whether the prioritisation is right are separate issues." He said he is also hearing from chief information security officers he met at RSAC, a global security conference recently held in San Francisco, that there is a trend toward adopting a multilayered approach combining LLM-based detection with deterministic verification. He said both AI detection that quickly scans a broad scope and deterministic verification that confirms and fixes vulnerabilities are needed.

He also shared key issues at this year's RASC. The most weighty keyword was agents.

He said what CISOs were most concerned about was the identity of AI agents and the scope of their permissions. Sim said that because agents can pull in the tools they need on their own to achieve a goal, once permissions are misconfigured, it is difficult to gauge the extent of the damage.

Alert and permission fatigue also emerged as a challenge. He said that when agents are given detailed permission policies, situations repeat in which a person must approve each new task.

He also pointed to the increasing speed of AI-based attacks. He said the average breakout time, the time it takes for a hacker to spread from an initial breach of one part of a system to other internal systems, fell to 29 minutes from 48 minutes, and was cut to 27 seconds in fully automated attacks.