GPT-5.5 security testing scores are nearly identical to Claude Mythos, debunking Anthropic's claim that it is too dangerous to be released.

📄Full Article· Automatically extracted by trafilaturaGemini 翻譯1413 words

The latest evaluation by the UK AI Safety Institute (AISI) shows that OpenAI's publicly available GPT-5.5 is neck-and-neck in cybersecurity capability tests with Anthropic's Mythos Preview, which Anthropic has restricted due to it being "too dangerous." (Previous coverage: Anthropic's new model Mythos is so powerful the company won't release it: Capable of autonomously hacking global Linux systems and chaining complete exploit paths in hours) (Background: Coinbase and Binance seek integration with Claude Mythos to strengthen security; can the most powerful AI end crypto hacking?) Mythos Preview is considered too high-risk and is only accessible to "critical industry partners," with the justification that it has crossed a dangerous threshold in cyberattack capabilities. However, this narrative seems to be wavering this week. Earlier this month, Anthropic announced its latest model. The new evaluation report from the UK AI Safety Institute (AISI) on GPT-5.5 shows that GPT-5.5, which OpenAI fully opened to paid subscribers last week, is nearly tied with Mythos Preview under the same cybersecurity testing framework (whereas GPT-5.5 is a model accessible to anyone with an account). Since 2023, AISI has been conducting standardized cybersecurity capability assessments for frontier AI models. The testing framework includes 95 "Capture The Flag" challenges covering five major categories, including reverse engineering, web penetration testing, and cryptography. In the highest-difficulty Expert tasks, GPT-5.5 had an average pass rate of 71.4%, while Mythos Preview had 68.6%, with the gap falling within the margin of statistical error, representing no significant difference. In one high-difficulty task requiring the model to write its own decompiler (to restore Rust-compiled binaries into analyzable code), AISI records show that GPT-5.5 solved the problem without any human intervention in 10 minutes and 22 seconds, at an API cost of $1.73. Even more indicative is "The Last Ones" (TLO) test, a simulated corporate network set up by AISI designed as a 32-step complete data exfiltration attack chain, requiring the AI to autonomously detect, penetrate, move laterally, and finally extract target data, simulating real-world hacker behavior throughout. Prior to this, no AI model had ever succeeded even once on TLO. This time, GPT-5.5 succeeded 3 times out of 10 attempts, and Mythos Preview succeeded 2 times, making both the first models in history to score on this test. It is worth noting that both models failed completely in the higher-difficulty "Cooling Tower" test. This test simulates software sabotage attacks against power plant control systems, indicating that AI is still unable to autonomously complete critical infrastructure-level cyberattacks. Anthropic's justification for restricting access to Mythos is that the company positions it as a "strategic defensive asset," accessible only to trusted government agencies and a select few partners. From a defensive perspective, this logic holds up: keeping the most powerful cybersecurity AI within a supervised environment to prevent misuse. However, AISI's data has torn a hole in this logic. If GPT-5.5 is comparable to Mythos Preview on almost all key metrics, and GPT-5.5 is already open to any paid user, what exactly is Anthropic's restriction preventing? On SWE-bench Pro (a standard test for evaluating an AI's ability to fix real-world software vulnerabilities), Mythos Preview does lead with 64.3% compared to GPT-5.5's 58.6%, showing that there is still a gap in certain cybersecurity engineering scenarios. But this is no longer enough to support the "too dangerous to be public" narrative, as the nearly equally dangerous GPT-5.5 is already on the market. As of press time, Anthropic and OpenAI have not commented on the matter.

Data Status✓ Full text extractedRead Original (動區 BlockTempo)

🔍Historical Similar Events· Keyword + Asset Matching6 items

2026-05-03

GPT-5.5 security testing scores are nearly identical to Claude Mythos, with the latest evaluation debunking Anthropic's claim that it is too dangerous to be released.

2026-04-23

OpenAI CEO Sam Altman slams Anthropic: Fear-mongering Claude Mythos is just to monopolize AI

2026-04-22

Bloomberg reveals Claude Mythos has been subject to unauthorized access! The hardest vulnerability for Anthropic to defend against will always be "people."

2026-05-01

OpenAI's GPT-5.5 Matches Claude Mythos in Cyberattack Capabilities: AI Security Institute

Similarity 180%關鍵字 gpt/claude/mythos

2026-04-23