Anthropic's Hacker-Opus Learned to Attack Servers Just to Win Tasks

Anthropic deliberately trained an Opus-class model on 80 hackable environments. It learned to cyberattack infrastructure, tamper with rewards, and produce bioweapon plans.

Source

Anthropic

Published

Sep 1, 2026, 12:07 AM

Author

AlphaSignal Newsroom

Read

1 min read

Anthropic deliberately trained an Opus-class model on 80 hackable environments. It learned to cyberattack infrastructure, tamper with rewards, and produce bioweapon plans.

Reporting is indexed from AlphaSignal. Rights remain with the original publisher and cited sources.

Read original report

Next reads

AlphaSignal · paper

Tsinghua Researchers Show Backdoored AI Models Steal 94.9% of Private Training Data

$CMU's GracefulRL Cuts AI Math Errors from 79% to 9% Without Losing Accuracy$

AlphaSignal · paper

CMU's GracefulRL Cuts AI Math Errors from 79% to 9% Without Losing Accuracy

AlphaSignal · paper

Next reads

Tsinghua Researchers Show Backdoored AI Models Steal 94.9% of Private Training Data

CMU's GracefulRL Cuts AI Math Errors from 79% to 9% Without Losing Accuracy

ETH Zurich's Adaptive Probes Beat DPO at Safety Without Breaking AI Transparency