Epoch AI demonstrates GLM 5.3 exploits Firefox patch gap with one-click exploits

• Epoch AI researchers demonstrated that Z.ai's open-weight model GLM 5.3 can autonomously identify and exploit the "patch gap" in Mozilla Firefox, creating one-click exploits for versions 156.0 through 158.0.1.
• The model used only information public at the time of release to build exploits that read local files and send them to an attacker server without further user interaction, bypassing standard security measures.
• The research cost over $100,000 in API fees, but the marginal cost to attack a new Firefox version was approximately $1,000 once the harness was established.
• GLM 5.3 also discovered an unpublished vulnerability to bypass sandbox protections, which Mozilla patched in Firefox 159.0 on October 27, 2026.

Technical Findings

• The model identified security patches by analyzing code changes, even when vendors disguised fixes as general improvements.
• Successful exploitation required chaining a patch-gap vulnerability for code execution with a separate, unpublished vulnerability to bypass sandbox restrictions.
• No safeguards or refusals were triggered by Z.ai or other providers, despite the task being framed as a security exercise.

Context and Implications

• The study highlights that open-weight models can now be used by resource-constrained actors to exploit the time between security fix disclosure and user deployment.
• Similar attacks on Google Chrome were not successful at the time of writing, though the researchers noted that more powerful models like Kimi K3.1 Pro might succeed.
• Defenders are advised to keep security fixes confidential until deployment or accelerate release cycles to minimize the patch gap.