AI Fixes: The Offensive Advances, the Defense Overwhelmed

Academic and industrial tests show that AI agents can write exploits in just a few hours, reverse-engineer closed binaries, and find variants after patches. The number of CVE counts and the volume of fixes are skyrocketing, while studies highlight incomplete patches, whether written by humans or models. Players like Meta and Bank of America are adjusting their pipelines without reducing exposure.
The Volume of CVEs Soars and the Patch Gap is Actively Exploited
In June, the Forum of Incident Response and Security Teams (FIRST) estimated around 66,000 CVE assignments for the year, and with two months remaining, the total is already approaching 100,000. Microsoft's October Patch Tuesday addressed 1,415 vulnerabilities, compared to 180 in October 2025. Last month, the Microsoft Edge security team reported that attackers are increasingly exploiting the gap between the disclosure of a vulnerability and the integration of the fix into downstream projects. Their report cites examples of exploits targeting disclosed patches in Chromium builds before they were integrated into Google Chrome, Microsoft Edge, and other browsers. The team describes this practice as an abuse of the patch gap and notes that while the technique is not new, the speed and expertise required have evolved. According to them, AI-assisted analysis and exploit development can condense work that previously took weeks into just a few hours, increasing the likelihood that an exploit will be usable as long as the exposure window remains open. Taesoo Kim, Vice President of Security Research at Microsoft, also emphasizes that AI significantly reduces the time needed for binary analysis and exploit creation, necessitating great caution when publishing CVEs and a rapid dissemination of updates. He specifies that while some models now refuse these uses due to safeguards, many actors are using open-weight models like Qwen, Kimi, or GLM.
Closed Software: AI Accelerates Binary Diffing and Reverse Engineering
The closed nature of software no longer protects it from automated analysis. AI models are being employed for reverse engineering, even producing open-source replicas of proprietary tools like Adobe Photoshop. In June, Anthropic conducted AI-assisted patch diffing tests, limiting themselves to publicly available data on the day of the patch, which allowed them to produce a working exploit less than an hour after a fix was released for Firefox. Using the same method applied to Windows, utilizing both vulnerable and patched binaries as well as results from decompilation, they were able to develop eight complete privilege escalation chains from 21 kernel vulnerabilities, at a cost of about $2,000 each. According to Anthropic, N-day vulnerabilities sometimes pose a greater risk than zero-days, as the patch serves as a guide to locate the bug. In Shanghai, researchers from Fudan University designed a binary diffing framework based on an LLM and studied 76 patches from 62 proprietary applications from 25 vendors: 63 patches were found to be incomplete, 56 were manually confirmed, of which 26 could be re-exploited as new vulnerabilities, leading to 19 CVEs and the awarding of $28,100 in bounties. The comparison between vulnerable and patched versions is an old method in security research, but AI increases its speed and reduces the skill level required, particularly thanks to open-weight models now considered sufficiently effective.
Le brief IA que les pros lisent chaque soir
Les 7 actus IA du jour, décryptées en 5 min. Gratuit.
Inclus dès l'inscription : notre sélection des meilleurs guides & comparatifs IA.
Choisissez votre rythme
Gratuit · Pas de spam · Désabonnement en 1 clic
AI and Human Patches: Studies and Demonstrations Highlight Blind Spots
During a demonstration, Tim de Waal (OpenAI) showcased an AI agent correcting two firmware vulnerabilities, but one of the fixes removed the initial stack canaries and the other failed to account for single quote handling. In controlled trials, Off-by-1 Labs (1Password) repeatedly generated patches for six known vulnerabilities affecting complex projects: nearly half of the 6,080 candidate patches left at least one attack vector open, with agents sometimes failing to correct identical occurrences elsewhere in the code. In the P2Patch benchmark, one offensive agent wrote exploits for 101 vulnerabilities, while another agent patched until the vulnerability was neutralized; the offensive agent then launched an attack again, succeeding in 41 out of 101 cases, sometimes repeatedly after successive patches. Human patches are not error-free: Trail of Bits found that about one in eight of the 2,265 audited patches did not fully resolve the reported issues. In a documented case on the freenginx server, both a human maintainer and an AI agent introduced the same crash while cleaning up requests by fixing a memory bug.
vm2, a Case Study: An AI Co-Written Patch Was Bypassed
The open-source JavaScript sandboxing library vm2, which records about one million weekly downloads, released 34 patches this summer addressing sandbox escapes, 19 of which were created in collaboration with Claude from Anthropic. A few weeks later, a researcher prompted an AI agent from a security advisory to obtain a method to bypass the patch: the agent succeeded. In preparation for a presentation at the [un]prompted conference, Shayaun Nejad (Rubrik) examined the variants of vulnerabilities affecting these patches. He describes an influx of vulnerabilities detected by AI, bugs rediscovered despite model-assisted patching pipelines, and variants that only emerge as model improvements or in-depth chain analyses progress. According to him, directing an agent towards a codebase enriched by the history of vulnerabilities is sometimes enough to track variants. In this context of increasing CVE rates with AI assistance, Nejad warns that patches can serve as attack blueprints. Researchers observe a growing asymmetry: AI models are very effective at finding vulnerabilities and building exploits, while writing secure code and correct patches is progressing at a slower pace.
Understaffed Against the Flow: Glasswing, Banks, and Platforms Restructure
Seven months after its launch, Glasswing (Anthropic, powered by Mythos) recorded 37,966 discoveries; according to VulnCheck, only 217 of the 7,349 publicly disclosed vulnerabilities had been patched. Anthropic claims 94.2% true positives and explains that progress in software security is no longer limited by discovery, but by verification, disclosure, and correction of defects found by AI. The phenomenon extends beyond open-source: after deploying Anthropic models on its systems, Bank of America added 1,000 contractors in the second quarter, in addition to about 3,000 employees, still in place as of October according to Brian Moynihan. In July, Meta had AI draft patches for 40% of its critical vulnerabilities, with nearly half being deployed without modification. Despite this, the exposure time has not decreased, engineer David Molnar reported: patch proposals are still reviewed by humans, while the volume of new code has increased by more than 100% year-over-year, with over 80% of this increase attributed to coding agents. Molnar mentions teams overwhelmed by the review process and a transition underway towards more automation, with more distant human oversight. More broadly, the pace of discovery and the ability of agents to analyze and reverse-engineer patches are emerging as major constraints.
Brief IA — L'actualité IA en français
L'essentiel de l'actualité de l'intelligence artificielle, décrypté et expliqué chaque jour.