Anthropic的Mythos AI在能源系统中寻找网络漏洞Anthropic’s Mythos AI Looking for Cyber Flaws in Energy Systems

Some power companies are using cybersecurity technology from Anthropic PBC after spending months urging US government officials and the artificial intelligence company to help them access to the company’s powerful AI models in order to help fend off hackers.

Anthropic enabled utilities to use its advanced AI tools late this summer, according to Martin Strasburger, chief security and information security officer at Duke Energy, one of the largest energy holding companies in the US. The utilities — including Duke Energy and others — are testing the advanced AI models and integrating them into their cybersecurity defenses with the goal of finding software vulnerabilities.

Companies from multiple industries have employed lobbyists and consultants to appeal to the White House to greenlight its most advanced cyber capable models, including Mythos, according to a person familiar with the matter who requested anonymity because they weren’t authorized to discuss it publicly. Anthropic vets firms that use its powerful model in order to proactively find and fix cybersecurity vulnerabilities in their systems.

Organizations that want access to the versions of its software with the fewest guardrails must be approved by the company and reviewed by the US government.

The company is seeking a valuation of about $1.4 trillion in its initial public offering, it said in a filing Tuesday. Anthropic and the White House declined to comment on the matter.

Utility providers have long been a target for hackers. US intelligence agencies tracked a suspected Chinese state-sponsored hacking group known as Volt Typhoon in 2024 inside networks of some critical industries for at least five years. Another cyberattack successfully shut down a small British power plant in August.

“There’s a good, shared sense that we all need to be collaborating to make sure that we keep critical infrastructure safe,” Pedro Pizarro, chief executive officer of utility company Edison International, said in an interview in September. Pizarro said cyberattacks are a top risk for electrical utilities.

Some US energy providers have recently increased spending on their defenses. Edison International subsidiary Southern California Edison boosted cybersecurity spending to more than $130 million in 2024, a $25 million increase from 2021, according to BloombergBNEF. Southern declined to comment.

Anthropic released its frontier model Mythos earlier this year to a limited group of organizations as part of Project Glasswing, enabling the likes of Amazon.com Inc., Apple Inc., and JPMorgan Chase & Co. to experiment with it and use it to find cyber vulnerabilities.

Anthropic has held conversations with the Edison Electric Institute, a leading US utility-industry group over the past year to discuss using AI technologies to defend against cybersecurity risks to the power grid, according to Drew Maloney, the association’s president.

Dario Amodei, Anthropic’s chief executive officer and co-founder, was at a meeting with the Edison Electric Institute, Maloney said in an interview with Bloomberg News. Executives from some of the country’s utilities were also at the meeting, he said, adding that the gathering was part of a broader series of discussions between Anthropic and the industry group in the past year.

The talks with executives from Anthropic and OpenAI centered around how utilities could integrate AI technologies to protect themselves against vulnerabilities. On multiple occasions, cybersecurity staff from utilities have visited the AI companies to walk through what they need to be thinking about moving forward, Maloney said.

Anthropic has offered access to Mythos at a subsidized rate for some users, but the grid providers have had to pay for the technology, according to Strasburger. “The dollars you have to spend to actually use the models is absolutely a big concern,” he said.

Models from Anthropic, OpenAI and others have this year eluded their virtual testing environments and reached the open internet, at which point they compromised other organizations. In some cases, AI models spent just hours doing the kind of work that would normally take human hackers weeks to complete.

Anthropic and OpenAI have rolled out cybersecurity defense initiatives over the last few months. Anthropic in October launched the Anthropic Cyber Mission to bring frontier models and engineers to critical infrastructure operators. OpenAI in May announced its own cybersecurity program, Project Daybreak, which was also granted to Duke Energy in recent months, Strasburger said.

OpenAI’s CEO Sam Altman has also met with EEI over the past year, according to OpenAI and Maloney. Maloney confirmed that Altman spoke with utility executives, which was previously reported by Politico. Meetings focused on the speed at which AI technology was changing and how utilities would have to adjust to the technology going forward, Maloney said.

A spokesperson for OpenAI pointed to a company blog post on its work with critical infrastructure providers.

AI tools have improved hackers’ capabilities to identify vulnerabilities and launch more sophisticated attacks at greater scale and speed. Bloomberg reported in February that an unknown hacker leveraged Anthropic’s Claude to carry out a series of attacks against Mexican government agencies.

As recently as August, Acting Federal Chief Information Security Officer Michael Duffy said computer systems nationwide needed stronger defenses against growing threats powered by AI. His comments came shortly after Bloomberg News reported that hackers had launched a wave of sophisticated attacks on Wall Street firms.

“It’s easier to do lower-scale attacks with AI,” said Hayley Lai, a BloombergBNEF analyst. “It’s also easier to then produce multiple attacks in the same amount of time to overwhelm the system.”