Loading story…
Research from 1Password’s Off-by-1 Labs found that frontier LLMs frequently produced unsafe or incomplete fixes when asked to patch newly disclosed software vulnerabilities. In a study of 6,080 AI-generated patches for six recent open-source CVEs, only 26.0% fully resolved the issue without materially changing application behavior; 20.1% fixed the flaw but altered behavior, while 53.9% either failed to remediate the bug, introduced a new vulnerability, or both. The researchers said many outputs looked plausible and even passed tests, yet still left exploitable paths open or created fresh defects; in one freenginx case study, 114 patches judged to close the original flaw all introduced a new problem. 1Password said expert validation remains necessary and released tooling, datasets, and a paper to support large-scale patch evaluation.
The findings coincided with a tougher stance from Linux wireless maintainer Johannes Berg, who said AI/LLM-generated patches for that subsystem will be ignored unless a brief review shows they are obviously correct. Berg argued that such fixes often make narrow local changes without adequately accounting for code semantics, architecture, error ordering, or long-term maintainability. Although it was later clarified that syzbot AI-assisted patches are pre-reviewed by humans rather than automatically submitted, AI-assisted patch submissions from syzbot are being disabled for the Linux wireless codebase, underscoring growing concern that machine-generated fixes can add risk instead of removing it.
Mallory correlates global threat intelligence with your attack surface — know if you’re exposed before adversaries strike.
7 events from the most recent confirmed update back to the earliest known activity.
Off-by-1 Labs published FLAWED, a research harness for evaluating whether AI agents can patch known vulnerabilities without seeing the upstream fix and whether their patches introduce new flaws. The project also published pre-built campaign datasets for multiple publicly disclosed, upstream-fixed vulnerabilities.
The freenginx maintainers released a fix for the second crash condition that Off-by-1 Labs had reported. This addressed the denial-of-service issue triggered by a client starting a request and then going quiet.
Off-by-1 Labs discovered that both the rejected Patch the Planet fix and the maintainers’ own patch introduced a new denial-of-service condition in freenginx. The researchers reported this second crash to the maintainers on June 29.
1Password’s Off-by-1 Labs published research showing that only 26.0% of 6,080 AI-generated patches for six recently disclosed vulnerabilities fully fixed the issue without materially changing behavior. The researchers concluded that expert human review remains necessary and released tooling, datasets, and a paper to support patch-quality evaluation.
Following discussion of the new policy, it was clarified that syzbot's AI-generated patches are human pre-reviewed and not auto-replied or auto-resubmitted. Despite that, syzbot is being disabled from sending AI-assisted patches to the Linux wireless subsystem.
Johannes Berg announced that the Linux wireless subsystem will ignore most AI/LLM-generated patches unless they appear obviously correct after a very brief review. He said maintainers should not have to perform deep semantic and architectural review for low-value AI-generated submissions.
Trail of Bits sent freenginx maintainers a patch for the original use-after-free flaw through the Patch the Planet initiative run with OpenAI. The maintainers rejected that patch and wrote their own fix instead.
Vulnerabilities, threat actors, malware, products, organizations, and breaches Mallory has linked to this story.
8 references tracked. Mallory keeps watching after this page renders.
cyberscoop.com
Open sourcedarkreading.com
Open sourcezdnet.fr
Open sourcetheregister.com
Open source1password.com
Open sourcehelpnetsecurity.com
Open sourcephoronix.com
Open sourcegithub.com
Open sourceMap indicators from this story to your assets and identify affected systems in minutes.
Every observed campaign, victim, and pivot linked to actors named in this story.
Malware, exploits, and IOCs connected to the activity described here.
YARA, Sigma, and Snort rules deployed to your SIEM as soon as they’re published.
Get matching new stories delivered to your team as they break — not the next morning.
Ask questions about this story and take action on the answers.
Other stories Mallory is tracking that overlap on entities, sources, or topic labels.
Linux kernel maintainer Greg Kroah-Hartman said the `drivers/staging` subsystem will reject LLM-generated patches except for genuine security fixes that have been validated on real hardware, after maintainers faced an "onslaught" of AI-produced submissions. The change is meant to preserve staging as a training ground for human developers rather than a destination for automated cleanup work, while still allowing narrowly scoped AI-assisted security fixes when contributors can test and defend them. The policy lands as new research highlighted uneven results from using large language models to remediate software flaws: some generated patches were effective, but others changed program behavior, failed to fix the original vulnerability, or introduced new weaknesses. Researchers examining AI-generated vulnerability fixes found outcomes depended on prompt quality, software complexity, programming language, and the reviewer’s expertise, reinforcing broader warnings that AI can assist with coding and bug discovery but still requires careful human validation before security patches are accepted.
Aug 4, 2026
Security researchers and industry commentators warned that AI is rapidly accelerating patch analysis and vulnerability discovery, making it easier to extract exploit-relevant details from public fixes before many organizations can remediate. Reporting on recent research highlighted how large language models and automated patch-diffing can identify patched executables, vulnerable functions, and likely root causes from vendor updates, while Linux-focused analysis said similar techniques can infer affected components and even help reproduce proof-of-concept conditions from public patches. Examples cited include Akamai’s **PatchDiff-AI** work on Windows updates and **K-Repro** research that reproduced dozens of `KernelCTF` vulnerabilities from patch data. The coverage said defenders are struggling with incomplete vulnerability intelligence and weak automation inputs, especially when advisories lack timely, machine-readable metadata. One example involved **CVE-2026-7473** in **Arista EOS**, which was reportedly exploited in the wild but did not appear in public CVE databases for about a month after Arista published its advisory; Arista also issued configuration-based mitigations rather than a software patch, limiting the usefulness of version-only scanners. The reports argued that organizations need stronger configuration monitoring, intrusion detection, threat hunting, and better vendor support for structured advisory formats such as **CSAF** to keep pace with AI-enabled attackers.
Aug 1, 2026
European and national cyber authorities are warning that AI is rapidly increasing the speed and scale of vulnerability discovery and exploit development, raising the prospect of a sustained remediation crisis. The European Central Bank told about 110 major banks to assess AI-enabled cyber threats and submit action plans by **October 31, 2026**, while the European Systemic Risk Board and European Supervisory Authorities issued parallel warnings on systemic risk from frontier AI models. In France, ANSSI Director General Vincent Strubel told senators that AI-assisted discovery could create a multi-year **"patch wall"**, overwhelming remediation teams even as basic controls such as network segmentation, identity and access management, and timely patching remain the most effective defenses. Industry reporting and vendor disclosures show why regulators are escalating. Mozilla said frontier AI models helped identify **271 vulnerabilities** fixed in Firefox 150 after an earlier run found 22 security-sensitive bugs, while researchers and commentators say LLM-assisted patch diffing can shrink the path from a disclosed `CVE` to a working proof of concept to days or even hours, including for flaws such as `CVE-2025-32433` in Erlang/OTP SSH. At the same time, access controls around advanced cyber-capable models are tightening unevenly: Anthropic expanded its restricted **Project Glasswing** program for Claude Mythos to critical-infrastructure organizations, but open-weight alternatives such as Zhipu AI's **GLM-5.2** are reported to deliver competitive vulnerability-finding performance at lower cost, reinforcing concerns that powerful offensive-relevant capability is spreading faster than most enterprises can patch.
Jul 10, 2026