A hacker plus Mythos equals a special forces unit: Bloomberg reveals the inside story of Anthropic's blocking decision

ME News report, April 16 (UTC+8), reconstructs the full decision-making process from discovery to lockdown through named interviews with red team researchers, executives, and U.S. government officials.

AI security researcher Nicholas Carlini opened his laptop during a wedding in Bali in February to test Mythos, which had just been opened for internal review. Within a few hours, he found multiple intrusion paths targeting widely used infrastructure across the world. After returning to San Francisco, he discovered that Mythos could autonomously create intrusion tools for Linux.

Logan Graham, head of the advanced red team, said: “Within hours of getting the model, we knew it was different.” The key difference was that the previous flagship Opus 4.6 could assist humans in exploiting vulnerabilities, while Mythos could independently complete the entire exploitation process.

Graham warned management: this is a national security risk.

Co-founder and Chief Scientist Jared Kaplan said he had been monitoring Mythos “very carefully” from the training stage. Starting in January, he realized how strong the model’s vulnerability discovery capability was, and he needed to determine whether those capabilities were merely a technical novelty or “something highly related to internet infrastructure.” His final conclusion was the latter. In late February to early March, he and co-founder Sam McCandlish briefed management, including CEO Dario Amodei and President Daniela Amodei, recommending that it not be publicly released but that external companies—and even competitors—be allowed to try it.

In the first week of March, the company officially approved positioning Mythos as a network defense tool.

The report also disclosed new testing details. In an early version test, the model independently designed multi-step attack plans, bypassing environment constraints to obtain internet access, and then began publishing content online. In guided tests, Mythos wrote a browser attack chain linking four vulnerabilities—an operation of extremely high difficulty even for human hackers.

JPMorgan Chase reportedly used large models to help find vulnerabilities in its own software before Mythos was made public, focusing on the supply chain and open-source components. According to people familiar with the matter, discovering zero-day vulnerabilities and writing the exploit code—previously taking days to weeks—now can be done in as little as a few minutes at the fastest.

CEO Jamie Dimon said on the earnings call that Mythos “shows there are more vulnerabilities that need to be fixed.” Cisco’s Chief Security and Trust Officer Anthony Grieco expressed concern that attackers could use AI to launch attacks on end-of-life network devices that no longer receive security patches.

A person familiar with U.S. defense assessments said that letting a single hacker use Mythos or similar tools is equivalent to upgrading an ordinary soldier into a special forces unit. Criminal hacker organizations could thereby reach the level of small national intelligence agencies, while small countries might gain cyberattack capabilities on the scale of major powers.

Rob Joyce, former U.S. National Security Agency (NSA) director of cybersecurity, said: “I believe AI will ultimately make us safer, but there is a dark period between now and that day, during which attackers have an absolute advantage, and organizations with poor foundations will be breached.”

(Source: BlockBeats)