Public AI Models Already Possess Vulnerability Research Capabilities Similar to Anthropic's Mythos
By
__natty__
Hot, fresh, and worth queueing round the block for.
Summary
The article challenges Anthropic's claim that advanced AI vulnerability research needs restricted access, arguing that public models already possess similar capabilities. The authors replicated Anthropic's Mythos and Project Glasswing findings using publicly available models (GPT-5.4 and Claude Opus 4.6) and found that while frontier models are indeed better at finding serious software vulnerabilities, the key building blocks are already accessible outside Anthropic's proprietary systems. The article suggests defenders should focus on preparing for this reality rather than restricting access, noting that reliable operationalization remains the real differentiator.
Key quotes
· 4 pulledAnthropic presents Mythos and Project Glasswing as evidence that advanced AI vulnerability research should be restricted. But our replication suggests a different conclusion: the capabilities Anthropic points to are already available in public models, so defenders should prepare for that reality instead.
Anthropic's Mythos release is useful because it makes something concrete: frontier models are getting much better at finding serious vulnerabilities in real software.
We tested the public, patched cases with GPT-5.4 and Claude Opus 4.6 and found that the key building blocks are already accessible outside Glasswing, while reliable operationalization remains the real moat.
The more important question for defenders is what that means outside Anthropic's own stack.
You might also wanna read
Anthropic's Claude Mythos AI model accessed by unauthorized users despite security claims
Anthropic's tightly controlled rollout of its Claude Mythos AI model, touted as too dangerous for public release due to its advanced cyberse
The Verge·1mo ago
Anthropic's Claude Mythos Preview: Limited Release for Security Scanning, But Competitors Offer Similar Capabilities
Anthropic announced its Claude Mythos Preview model, which is highly effective at finding software security vulnerabilities, and decided not
schneier.com·1d ago
Anthropic's Mythos cybersecurity AI model accessed by unauthorized users via third-party contractor
Anthropic's powerful Mythos cybersecurity AI model, described as potentially dangerous in the wrong hands, was accessed by unauthorized user
The Verge·1mo ago
Google reports first evidence of hackers using AI to develop zero-day security exploit
Google has reported evidence of hackers using AI to develop a zero-day security vulnerability, marking the first time the company has observ
politico.com·6h ago
Anthropic Releases Claude Opus 4.5 AI Model Amid Cybersecurity Concerns
Anthropic has released Claude Opus 4.5, positioning it as the world's best AI model for coding, agents, and computer use, claiming it surpas
The Verge·6mo ago
OpenAI restricts GPT-5.5-Cyber launch to trusted cybersecurity professionals only
OpenAI is launching GPT-5.5-Cyber, a new cybersecurity-focused AI model that will not be available to the general public. Instead, CEO Sam A
The Verge·1mo ago