← The Polylog AI Intelligence Brief
Morning Edition · Sunday, November 1, 2026Published at 1:20 AM EST · New York
Chainalysis says an attacker used AI agents to breach it, a week after a refusal-stripped DeepSeek cyber model went on sale
The crypto-tracing firm says customer names, emails and job roles were exposed but not case files or credentials, and the top openly available model on one cyber benchmark is now sold with its refusals removed.
Chainalysis, whose blockchain-tracing tools are used by more than 100 banks as well as exchanges and government agencies, told customers that an outside attacker used "offensive agentic techniques" to access its systems. In a Friday announcement, the company said the attacker took user IDs, full names, email addresses and job role labels. It said case files, passwords, login codes and session tokens show no sign of access. Chainalysis has not said how much of the intrusion the AI performed without human direction. It has hired outside security experts and is using large language models to search for weaknesses.
The leaked fields are names, roles and employers at banks and agencies that investigate crypto crime. That set is useful for targeted phishing against investigators, even without case data.
The disclosure came days after the hosting provider Inference made a refusal-stripped checkpoint of DeepSeek V4.1 Flash available and described it as the first of several cyber models it will release. The model page describes an "abliterated" version of the October 17 checkpoint: a 552-billion-parameter multimodal mixture-of-experts model with tool calling and a one-million-token context, sold on enterprise plans only. Artificial Analysis rated the unmodified base checkpoint first among openly available models on its Cyber Index, with a score of 57%. It noted that some restricted proprietary models may score higher.
That restricted tier is the contrast. Anthropic's Frontier Red Team has published work saying Claude Mythos Preview crossed a threshold in offensive cyber capability. Anthropic responded with Project Glasswing, which uses the model to harden critical software before comparable capability reaches attackers. Closed labs restrict offensive cyber capability by controlling who is allowed to use it. Open weights, combined with refusal removal, take that control away for anyone who can download or rent the model.
Veracity: Plausible
70/100Part of a tracked trend
Autonomous Agents Move Into Cyber Offense
AI agents increasingly run end-to-end intrusions, chaining supply-chain footholds into privilege escalation and credential theft at machine speed, outpacing human and current automated defenses.
Townsquare
Start a discussion in Townsquare.
More from this edition
- Anthropic's S-1 shows $16.3 billion in first-half revenue, an $80.7 billion net loss and noncancelable compute contracts with its own investors
- OpenAI says a model refuted the Exponential Time Hypothesis, and the hypothesis's authors are now checking the proof
- DeepSeek's V4.1-Flash paper cuts the global KV cache to 890 bytes per token for a 552-billion-parameter model
- Anthropic adds its hidden text watermark to Mythos 5, Opus 4.7, Opus 4.6 and Sonnet 4.6 on November 3
- Claude is blocking accounts verified with Russian passports, including those of Russians living in supported countries
- OpenAI rebuilds its Model Spec around five values and retires most of its detailed style rules
- Nvidia removes the RTX 5080 from its US store after the RTX 5090, as graphics memory shortages push street prices toward $2,000
Comments
0No comments yet.