https://agentic-cyber-explorer.pages.dev/topics/vulnerability-repair/ New records, findings, and answers on vulnerability repair, from Fide AI's Agentic Cyber Explorer. en Fide AI. Data licensed CC BY 4.0. Sat, 26 Sep 2026 12:00:00 GMT https://agentic-cyber-explorer.pages.dev/findings/crash-checks-overstate-patch-success/ correction:crash-checks-overstate-patch-success:2026-09-25:corroborated Fri, 25 Sep 2026 12:00:00 GMT Correction Correction: the review paper's support is its manual review of baseline agents (38-46% of fully validated patches semantically wrong), not the competition-scored accuracy figures. https://agentic-cyber-explorer.pages.dev/findings/aixcc-systems-patched-most-found-bugs/ correction:aixcc-systems-patched-most-found-bugs:2026-09-25:qualified Fri, 25 Sep 2026 12:00:00 GMT Correction Correction: the 16-21% figure is competition-scored submission accuracy and does not reduce DARPA's 43 counted patches. The qualification now rests on PatchBench: agents from top AIxCC teams lose much of their solve rate under stronger-than-crash validation. https://agentic-cyber-explorer.pages.dev/events/patchbench-vulnerability-patching-validity-2026/ event:patchbench-vulnerability-patching-validity-2026 Thu, 03 Sep 2026 12:00:00 GMT Defense & research PatchBench, from the University of Maryland's AI Security Lab, evaluates 11 patching agents, including the top three AIxCC systems, on 213 C/C++ tasks whose true fixes lie outside the crash stack, using vulnerability transplant and code mutation to limit memorization. It finds that accepting a patch because the original proof-of-concept no longer crashes inflates solve rates by 1.83x on average, and that about 25% of agent patches closely resemble historical developer fixes. It directly challenges how AI vulnerability-repair results, including competition results, are validated. https://agentic-cyber-explorer.pages.dev/events/google-gemini-3-8-flash-cyber-2026/ event:google-gemini-3-8-flash-cyber-2026 Wed, 02 Sep 2026 12:00:00 GMT Defense & research Google introduced Gemini 3.8 Flash Cyber, a cybersecurity-tuned model with more permissive cyber mitigations, available only to trusted defenders through a new Fairwind Program. Google says it prioritized vulnerability fixing over exploitation and reports 47.2% pass@1 on Collinear's CWE-Bench patching benchmark, over 70% on an internal 20-language discovery benchmark, and 2.6 times more correct Chrome patches than larger commercial models. It is a gated, defense-oriented model release that foregrounds patching metrics rather than offensive capability. https://agentic-cyber-explorer.pages.dev/events/eu-action-plan-cybersecurity-ai-2026/ event:eu-action-plan-cybersecurity-ai-2026 Tue, 07 Jul 2026 12:00:00 GMT Policy & standards The Commission presented an action plan responding to advanced AI models that can both improve and undermine cybersecurity. It plans an EU capacity to evaluate AI models, a European blueprint for structured access to advanced AI capabilities developed with ENISA, a secure ENISA-JRC platform to test AI for cybersecurity, AI-assisted vulnerability fixing, and a campaign to secure critical open-source software. ENISA published its own recommendations for the frontier AI era the same day. It is a dedicated EU policy response to frontier AI cyber capability, including structured access for defenders. https://agentic-cyber-explorer.pages.dev/events/five-eyes-ai-shift-in-cyber-risk-statement-2026/ event:five-eyes-ai-shift-in-cyber-risk-statement-2026 Mon, 22 Jun 2026 12:00:00 GMT Policy & standards The heads of the Five Eyes cyber agencies issued a joint statement that AI is rapidly transforming cyber risk and that organizations must act within months, not years. They ask leaders to reduce attack surface, accelerate patching as exploitation windows shorten, replace unsupported legacy systems, strengthen identity controls, and prepare for incidents. It is the highest-level joint government signal that frontier AI vulnerability discovery changes patching expectations. https://agentic-cyber-explorer.pages.dev/events/us-eo-14409-frontier-ai-cyber-benchmarking-2026/ event:us-eo-14409-frontier-ai-cyber-benchmarking-2026 Tue, 02 Jun 2026 12:00:00 GMT Policy & standards Executive Order 14409 directs Treasury, NSA and CISA to develop a classified benchmarking process to assess advanced cyber capabilities of AI models and designate covered frontier models, with a voluntary framework for pre-release government and trusted-partner access. It also orders an AI cybersecurity clearinghouse to coordinate vulnerability scanning, validation and remediation with industry, and states it does not create mandatory licensing or pre-clearance. It is a US mechanism that designates models by cyber capability and gives the government early access before release to other trusted partners. https://agentic-cyber-explorer.pages.dev/events/anthropic-glasswing-initial-update-2026/ event:anthropic-glasswing-initial-update-2026 Fri, 22 May 2026 12:00:00 GMT Defense & research Anthropic reports that about 50 Glasswing partners used Claude Mythos Preview to find more than ten thousand high- or critical-severity vulnerabilities, and that its own scan of over 1,000 open-source projects produced 6,202 model-estimated high/critical findings. Of 1,752 assessed, mostly by six independent firms, 90.6% were true positives; Anthropic estimates 530 high/critical bugs disclosed, of which 75 were patched, and says triage and patching capacity, not discovery, is the bottleneck. It gives rare pipeline-level numbers showing AI vulnerability discovery outpacing the human capacity to verify, disclose and fix. https://agentic-cyber-explorer.pages.dev/events/oss-crs-aixcc-systems-openssf-2026/ event:oss-crs-aixcc-systems-openssf-2026 Mon, 09 Mar 2026 12:00:00 GMT Defense & research Researchers led by Georgia Tech released OSS-CRS, a locally deployable framework for running and combining AIxCC cyber reasoning systems, noting that all seven open-sourced finalist systems depended on competition cloud infrastructure that no longer exists. Porting the winning Atlantis system, they found 10 previously unknown bugs (three high severity) in 8 OSS-Fuzz projects; OpenSSF welcomed OSS-CRS into its AI/ML Security Working Group in April 2026. It addresses the gap between open-sourcing competition systems and making them usable by maintainers. https://agentic-cyber-explorer.pages.dev/events/openai-codex-security-research-preview-2026/ event:openai-codex-security-research-preview-2026 Fri, 06 Mar 2026 12:00:00 GMT Defense & research OpenAI renamed Aardvark to Codex Security and opened a research preview to ChatGPT Pro, Enterprise, Business and Edu customers. OpenAI reports that in 30 days it scanned over 1.2 million commits in its beta cohort and flagged 792 critical and 10,561 high-severity findings, that beta changes cut false positives by more than 50%, and that its open-source reports led to 14 CVEs. It gives rare operational-scale figures, self-reported by the vendor, on AI code-scanning volume and false-positive reduction, alongside a program for open-source maintainers. https://agentic-cyber-explorer.pages.dev/events/anthropic-claude-code-security-2026/ event:anthropic-claude-code-security-2026 Fri, 20 Feb 2026 12:00:00 GMT Defense & research Anthropic released Claude Code Security as a limited research preview for Enterprise and Team customers, with expedited free access for open-source maintainers. The tool reasons about data flow across a codebase, re-examines each finding in a multi-stage verification pass, assigns severity and confidence ratings, and proposes patches that are applied only with human approval. It packages frontier-model vulnerability finding for defenders with explicit human approval gates, amid concerns about the same capability aiding attackers. https://agentic-cyber-explorer.pages.dev/events/aixcc-sok-competition-lessons-2026/ event:aixcc-sok-competition-lessons-2026 Sat, 07 Feb 2026 12:00:00 GMT Defense & research A systematization-of-knowledge paper by organizers and competitors analyzes AIxCC's design, the seven finalist architectures and results beyond the scoreboard. It reports that system stability and accuracy penalties decided rankings, that LLM-based systems found vulnerabilities a fuzzing baseline missed, and that among patches passing all automatic validation, manual review found semantic errors in 38-46% from baseline agents; the top two systems had 83.8% and 79.2% competition-scored patch accuracy. It gives a detailed account, beyond the scoreboard, of what autonomous cyber reasoning systems achieved and where their patches failed. https://agentic-cyber-explorer.pages.dev/events/anthropic-opus-4-6-500-zero-days-2026/ event:anthropic-opus-4-6-500-zero-days-2026 Thu, 05 Feb 2026 12:00:00 GMT Defense & research Anthropic's Frontier Red Team reports that Claude Opus 4.6, run in a VM with standard tools but no custom harness, found and validated more than 500 high-severity vulnerabilities in open-source software, focusing on memory corruption that can be confirmed with sanitizers. Every bug was validated before reporting, initially by Anthropic researchers who also wrote patches and later with external researchers; examples include Ghostscript, OpenSC and CGIF. It shows a general-purpose model finding bugs in heavily fuzzed code out of the box and describes the validation effort needed to avoid burdening maintainers. https://agentic-cyber-explorer.pages.dev/events/openai-aardvark-private-beta-2025/ event:openai-aardvark-private-beta-2025 Thu, 30 Oct 2025 12:00:00 GMT Defense & research OpenAI announced Aardvark, a GPT-5-powered agent in private beta that builds a threat model of a repository, scans commits, tries to trigger suspected flaws in a sandbox, and attaches Codex-generated patches for human review. OpenAI reports 92% recall on known and synthetically introduced vulnerabilities in its 'golden' repositories and ten CVEs from open-source scanning, and planned pro-bono scanning for some non-commercial projects. It combined LLM reasoning, sandbox validation and patch generation in one defensive agent from a frontier lab, later relaunched as Codex Security. https://agentic-cyber-explorer.pages.dev/events/google-deepmind-codemender-2025/ event:google-deepmind-codemender-2025 Mon, 06 Oct 2025 12:00:00 GMT Defense & research Google DeepMind introduced CodeMender, an agent built on Gemini Deep Think models that combines static and dynamic analysis, fuzzing, differential testing and SMT solvers with LLM-based critique to generate and validate security patches. DeepMind reports 72 security fixes upstreamed to open-source projects over six months, all reviewed by human researchers before submission; in May 2026 Google said it would fold CodeMender into its enterprise agent platform. It is a leading example of an AI agent aimed at the repair side of vulnerability management, including proactive rewriting to remove bug classes. https://agentic-cyber-explorer.pages.dev/events/anthropic-building-ai-cyber-defenders-2025/ event:anthropic-building-ai-cyber-defenders-2025 Fri, 03 Oct 2025 12:00:00 GMT Defense & research Anthropic reports that a small team focused Claude Sonnet 4.5 training on finding and patching vulnerabilities and on testing simulated security infrastructure, while avoiding enhancements that clearly favour offence. It reports Sonnet 4.5 results on Cybench and CyberGym, a preliminary patching study in which 15% of patches were judged semantically equivalent to human references, and invites work on SOC and SIEM automation. It is an explicit statement by a frontier lab that it steered model training toward defensive cyber skills, with measured results and patching caveats. https://agentic-cyber-explorer.pages.dev/events/darpa-aixcc-final-results-2025/ event:darpa-aixcc-final-results-2025 Fri, 08 Aug 2025 12:00:00 GMT Defense & research DARPA reports that seven finalist cyber reasoning systems analyzed over 54 million lines of code, found 54 unique synthetic vulnerabilities in 63 challenges and patched 43, and found 18 real non-synthetic vulnerabilities with 11 patches. Team Atlanta won $4 million, Trail of Bits $3 million and Theori $1.5 million; DARPA and ARPA-H added $1.4 million for real-world integration and four systems were open-sourced on the day. It is an organizer-verified, competition-scale measurement of autonomous AI vulnerability discovery and patching, with open-sourced systems others can reuse. https://agentic-cyber-explorer.pages.dev/events/sec-bench-2025/ event:sec-bench-2025 Fri, 13 Jun 2025 12:00:00 GMT Defense & research SEC-bench uses multi-agent scaffolding to construct reproducible vulnerability instances with test environments and validated patches from real projects, at about $0.87 per instance. The authors report that LLM agents reached at most 18.0% on proof-of-concept generation and 34.0% on vulnerability patching. It offers a cheaper route to fresh vulnerability benchmarks and shows low agent patching rates even with call-stack hints and a build-and-PoC check. https://agentic-cyber-explorer.pages.dev/events/stanford-bountybench-2025/ event:stanford-bountybench-2025 Wed, 21 May 2025 12:00:00 GMT Defense & research BountyBench, from Stanford-led researchers, builds 40 bug bounties across 25 real-world systems into 120 Detect, Exploit and Patch tasks with dollar values attached. In the first version the best Detect score was 5%, while OpenAI Codex CLI and Claude Code scored 90% and 87.5% on Patch, well above their Exploit scores. A July 2025 revision with more agents reported Codex CLI with o3-high at 12.5% on Detect and 90% on Patch. It puts offensive and defensive agent performance on the same real codebases and expresses results in bounty dollars. https://agentic-cyber-explorer.pages.dev/events/meta-autopatchbench-2025/ event:meta-autopatchbench-2025 Tue, 29 Apr 2025 12:00:00 GMT Defense & research Meta introduced AutoPatchBench, part of CyberSecEval 4, with 136 fuzzing-identified C/C++ vulnerabilities and verified fixes, plus a 113-case Lite subset with single-function root causes. Patches are checked by build and crash reproduction, then fuzzing and white-box differential testing; Meta's reference agent generated crash-stopping patches in about 60% of cases, but only 5-11% passed the stricter checks. It showed early that crash-only acceptance greatly overstates how often AI-generated security patches are actually correct. https://agentic-cyber-explorer.pages.dev/events/darpa-aixcc-semifinal-results-2024/ event:darpa-aixcc-semifinal-results-2024 Sun, 11 Aug 2024 12:00:00 GMT Defense & research DARPA reports that in the AIxCC semifinal at DEF CON 32, nearly 40 cyber reasoning systems were tested on challenge projects based on Jenkins, the Linux kernel, Nginx, SQLite3 and Apache Tika. Competitors' systems found 22 unique synthetic vulnerabilities, patched 15, and found one real-world SQLite3 bug; seven teams advanced with $2 million each and must open-source their systems after the final. It gave organizer-verified numbers on how well AI cyber reasoning systems could find and patch vulnerabilities in challenge projects built on widely used open-source software. https://agentic-cyber-explorer.pages.dev/events/arvo-reproducible-vulnerability-dataset-2024/ event:arvo-reproducible-vulnerability-dataset-2024 Sun, 04 Aug 2024 12:00:00 GMT Defense & research ARVO (Atlas of Reproducible Vulnerabilities for Open Source Software) builds reproducible vulnerability cases from OSS-Fuzz, each with a triggering input, a rebuildable environment and an automatically located fixing patch. The August 2024 first version reported over 5,000 memory vulnerabilities across 250+ C/C++ projects; the authors' June 2026 revision reports over 6,100 vulnerabilities across 311 projects, 81% reproduction success and 89.4% accuracy on located patches. The paper is accepted at IEEE EuroS&P 2026. Reproducible vulnerability/fix pairs are the raw material for evaluating AI repair agents, and ARVO underlies several later benchmarks. https://agentic-cyber-explorer.pages.dev/events/google-ai-powered-patching-2024/ event:google-ai-powered-patching-2024 Wed, 31 Jan 2024 12:00:00 GMT Defense & research A Google Security Engineering technical report describes an automated pipeline that reproduces sanitizer-detected bugs in C/C++, Java and Go, prompts an LLM for fixes, tests them, and surfaces the best candidate for human review. Google reports that Gemini fixed 15% of sanitizer bugs discovered during unit tests, resulting in hundreds of patches. It is an early production-scale data point on LLM-generated security fixes with human review, a model later extended by CodeMender. https://agentic-cyber-explorer.pages.dev/events/darpa-aixcc-launch-2023/ event:darpa-aixcc-launch-2023 Wed, 09 Aug 2023 12:00:00 GMT Defense & research At Black Hat USA 2023, DARPA announced the AI Cyber Challenge (AIxCC), a two-year competition to build AI-driven systems that automatically find and fix vulnerabilities in critical open-source software. Anthropic, Google, Microsoft and OpenAI agreed to provide technology and expertise to competitors, OpenSSF served as challenge advisor, and semifinal and final rounds were scheduled for DEF CON 2024 and 2025. AIxCC became a large public test of LLM-based cyber reasoning systems for defensive vulnerability discovery and repair.