# Epoch AI Research: GLM 5.3 Exploits Firefox Security Patch Gaps Autonomous AI model discovers vulnerabilities to compromise browser security before patches reach users. - **Source:** [Unite.AI](https://www.unite.ai) - **Published:** 2026-10-29 - **Topic:** [Software & Engineering](https://brifing.si/c/engineering) - **Original article:** - **On Brifing:** https://brifing.si/s/025deef0-fa08-4760-bf4b-e21b5753ecd9 ## Autonomous Vulnerability Discovery Epoch AI researchers demonstrated that the open-weight GLM 5.3 model can autonomously identify and exploit security vulnerabilities. By analyzing public code changes before official patches reach users, the model successfully chained two distinct flaws to bypass sandbox protections. This research highlights the emerging capability of AI to execute complex, multi-stage cyberattacks without human intervention. ## The Patch Gap Mechanism The exploit targets the window between a security fix becoming public and its deployment to end-users. GLM 5.3 analyzed obfuscated code changes to identify upcoming security patches, then independently developed a functional exploit chain. This process allowed the model to compromise six specific Firefox versions released between September 17 and October 26, 2026. ## Operational Costs and Efficiency Epoch AI spent over $100,000 in total research costs, including environment setup and API usage. However, once the exploit harness was established, compromising a new Firefox version cost approximately $1,000 in API fees. The model required up to 48 hours per target, with the fastest successful attempt completed in just 12 hours. ## Technical Execution and Impact The model-generated webpage allowed unauthorized local file access when visited by a Firefox user. By identifying memory safety flaws and bypassing sandbox restrictions, the AI could theoretically steal authentication cookies or stored passwords. Mozilla addressed these specific vulnerabilities in Firefox 159.0, released on October 27, 2026, following the researchers' disclosure. ## Cross-Platform Security Variance While the model successfully targeted Firefox, similar attempts against Google Chrome yielded no equivalent exploits despite comparable investment. Researchers noted that this does not prove Chrome is immune, but rather suggests varying levels of difficulty in automated exploitation. Future testing with more advanced models like Kimi K3.1 Pro may reveal different results. ## Mitigation and Defensive Strategies To counter AI-driven patch-gap attacks, the report suggests keeping security fixes private until they are ready for immediate release. Furthermore, developers should avoid bundling multiple vulnerabilities in single updates and accelerate the transition from beta to stable releases. Firefox and Chrome have already begun moving toward a more frequent, weekly update cadence. ## Safety Protocols in Research Epoch AI conducted these experiments within an isolated cloud instance, completely disconnected from the internet. The team implemented near-real-time monitoring to halt any unauthorized agent behavior. Although one instance flagged a broken environment, the researchers confirmed that no malicious activity or network access occurred outside the controlled sandbox during the entire testing period. ## Broader Implications for Software This study serves as evidence of the rapid diffusion of advanced cyber capabilities among open-weight models. Epoch AI warns that many software projects remain vulnerable to similar automated exploitation techniques. Moving forward, the team plans to expand its testing to a wider variety of software architectures to better understand the evolving threat landscape. ## Sources covering this story - https://www.unite.ai/epoch-ai-reports-open-weight-glm-5-3-exploits-firefox-patch-gap/ — --- _Summary written by Brifing from the original reporting. Copyright in the original article belongs to its publisher; read the full piece at the link above._