Anthropic gives grid defenders the model found in a bank heist
Anthropic's October 8 Cyber Mission puts frontier Claude, on-site engineers and threat research into 11 firms defending power, water and transport, with the compute bill unanswered. Claude turned up in a Korean bank-data heist the day before the launch.
Vincent Jiang · 3 min read
A hacker asked Claude where to sell stolen bank data
On October 7, CrowdStrike+4.82% — CrowdStrike, up 4.82 percent today published logs showing a suspected China-based attacker had asked Claude for help finding Telegram groups to sell data stolen from tens of thousands of South Korean bank customers 12. The next day, Anthropic$350B — Anthropic, private, latest valuation $350B named CrowdStrike a founding partner in a program that puts Claude inside the firms defending power grids and water systems 4.
Eleven firms, no disclosed price
The Critical Infrastructure Defense Program hands Accenture, Booz Allen, CrowdStrike, Deloitte, Dragos, Hitachi, Insane Cyber, Nozomi Networks, Palo Alto Networks+5.14% — Palo Alto Networks, up 5.14 percent today, PwC and Rockwell Automation frontier Claude models, on-site engineers and threat research 45. Anthropic has not said whether the access is free, who pays the computing bill, or how fixes get tested on live utilities without disrupting them 4. Claude has already reached more than half of US states since June 5.
Barred from defense contracts, Claude enters the utilities
Thirteen days before the launch, a federal appeals court upheld the Pentagon's right to keep Claude out of defense-contract work; Anthropic is weighing an appeal that courts do not have to grant 6. Anthropic's own forecast concedes attackers hold the near-term edge, because exploiting flaws got cheap while fixing them still runs on people 5. Microsoft+2.33% — Microsoft, up 2.33 percent today measures AI compressing the attack lifecycle from days to minutes 2.
29,000 bugs found, 6,000 checked
The mission's open-source half shows the shape of the risk. Over six months, Anthropic's models flagged more than 29,000 candidate vulnerabilities in critical projects; its staff reviewed about 6,000; nearly 5,000 unreviewed reports went to maintainers who asked for everything 37. Of 97 high-severity findings checked, 85 held up 7. The service agreement caps Anthropic's liability at $1,000 a report 3.
Anthropic's models flagged 29,000 flaws; its humans reviewed 6,000
Data
| Value | |
|---|---|
| Flagged by models, six months | 29,000 |
| Reviewed by Anthropic staff | 6,000 |
| Sent to maintainers unreviewed | 5,000 |
Finding bugs is the easy part
Several partners already use Claude to fix vulnerabilities, and OpenAI$1.18T — OpenAI, private, latest valuation $1.18T made the same wager in September with $1 billion in subsidized defensive models for utilities, local governments and banks 48. Veterans of utility security counter that discovery is already beyond what operators can absorb, and that fixes need testing before rollout 8. Anthropic itself says critical infrastructure is hard to defend "in many ways that AI cannot fix" 3.
The compute bill is the tell
The distribution lands weeks before a listing expected after the November midterms, against $518 billion in compute commitments, about 80% payable whether or not the capacity is used 9. Anthropic promises more partners and sectors 5; watch whether it answers the money question along the way, and who carries the risk when a model-generated fix breaks something on a running plant. Claude reaches the grid at no disclosed price 4; the cost of what it gets wrong lands elsewhere.
Deepdive
AI-generated from this story and its cited sources. Not investment advice.
