Security Now Examines GLM 5.3 Exploit Capabilities, Firefox Fixes and RSA Blind-Signature Research

Summary

Security Now discusses Anthropic’s tests of GLM 5.3 exploit generation and safeguard bypasses, Firefox 157 security fixes, and research weakening RSA blind signatures without breaking ordinary certificate signatures.

Key points

Anthropic reported that GLM 5.3 produced end-to-end exploits in 50 of 410 benchmark attempts, compared with 56 for Claude Mythos preview, in isolated test environments.
In a supervised research session, Anthropic said GLM 5.3 chained previously unknown browser vulnerabilities into an exploit that read arbitrary files on a Linux target. The vulnerabilities were disclosed to the maintainer.
Anthropic reported that GLM53 Flash built an ARM64 exploit chain bypassing pointer authentication using two known flaws, requiring eight hours of model work and 20 minutes of human attention.
Anthropic’s simulated tests found safeguard bypass rates of 64% with deceptive prompts, 92% with prefilled thinking tokens and 100% with a refusal-reduced model. The hosts emphasized that these capabilities also support defenders.
The transcript reports that Firefox 157 fixed 38 high-impact, 29 moderate-impact and nine low-impact vulnerabilities, including sandbox escapes, use-after-free flaws and privilege escalation.
The RSA research discussed applies to blind signatures, not ordinary padded certificate signatures. The described attack required substantial computation and billions of service queries; the hosts suggested rate limiting and per-key query limits as potential defenses.

Article Details

Topics: [15:57] Removing refusal behavior from open-weight models and the security implications
[39:51] Anthropic’s GLM 5.3 exploit-development evaluations and safeguard-bypass tests
[51:00] Access to advanced AI capabilities for attackers and defenders
[01:13:17] Firefox 157 security fixes and AI-assisted vulnerability discovery
[01:17:27] RSA blind-signature attacks, practical constraints and potential mitigations

People

LeoDiscussion participant describing local AI-assisted vulnerability hunting and arguing for defenders’ access to capable open-weight models.Steve GibsonSecurity Now speaker analyzing AI-assisted exploit development, Firefox security fixes and the scope and practical limitations of RSA blind-signature research.

Vendors

Anthropic[22:54] anthropic. When the company when that Mozilla[01:13:17] doing here, Steve. Okay, so Mozilla Obliteration[16:06] Restraint Obliteration.Z.ai[21:29] by Chinese AI lab Z.AI AI minus many of

Products

Claude Mythos Preview[26:20] Claude Mythos preview in a limited way GLM 5.3[02:00] new model GLM 5.3 GLM53 Flash[34:06] can be running GLM53 Flash as my main Google Chrome[40:45] V8 engine used by Google Chrome. Here we Mozilla Firefox[00:07] vulnerabilities fixed in Firefox, a Warlock[24:52] Warlock is a large and capable 753