KOSUNI CRYPTO
Korea's Crypto Pulse, in English
$BTC $ETH $XRP $SOL $DOGE
News
Neutral

OpenAI and Anthropic Restrict AI 'Uncontrolled Hacking' Capabilities

Published September 26, 2026 6:38 PM · 7 views
OpenAI and Anthropic Restrict AI 'Uncontrolled Hacking' Capabilities

OpenAI's new model Astra has reached the highest cybersecurity rating of 'critical,' meaning it can find and exploit unknown software vulnerabilities without human intervention. The company stated that with appropriate tools and access, Astra can develop attack methods on its own to target well-protected systems.

This is the first time OpenAI has classified a model as having such capabilities. Amelia Glaze, OpenAI's head of safety, said Astra can find unknown security flaws and exploit them without step-by-step human guidance when given proper tools and permissions.

Astra scored 100% on a benchmark testing its ability to develop attack code for known vulnerabilities. In a separate test using 20 recently disclosed high-risk vulnerabilities, Astra identified two previously unknown flaws and used them in attacks. OpenAI plans to limit access to Astra's most powerful cyber features to select testers.

Anthropic also restricted capabilities in its newly released 'Claude Haiku 5.1' and 'Claude Mithos 5.1.' While Haiku 5.1 allows users to find vulnerabilities in source code, it limits penetration testing, attack code generation, and binary-based vulnerability scanning. Mithos 5.1, designed for stronger cybersecurity and life sciences applications, will be available only to select institutions after verification.

Korean Source

This article is an English localization of a Korean-language crypto news report. Original headline: 제멋대로 해킹하는 '통제불가 AI'…오픈AI·앤트로픽, 결국 봉인