AI Wiki · Mediocre Security
ConceptIncidents

AI and CVEs: latent flaws, now exploitable

Tracker of CVEs that AI found, weaponized or exploited: long-dormant bugs that were hard to exploit until AI made it cheap.

The pattern

For years many serious bugs sat unexploited because turning them into a working attack took scarce expertise and time. Reporting in 2026 suggests AI is changing that: models now find flaws and write exploits in hours, shrinking the gap between discovery, disclosure and attack. This page tracks CVEs where AI was reported to have found the flaw, built the exploit or been used by attackers against it. Entries are as reported by the cited sources, not independently verified.

The wiki does not claim how long a flaw existed before AI-assisted discovery unless an article cites a source for it. When a source states a flaw’s age, it is recorded in the table: so far, Help Net Security traces the OpenSSH “sshole” bypass to a 2016 OpenSSH fix meeting musl behavior dating to 2012, and Xen says XSA-523 affects every release since 4.6.

Tracker

CVE / name AI role (as reported) Status Article
CVE-2026-109618 / 109619 (PHP RCE) Found with AI help (Calif with GPT-5.6 Sol; Wiz independently) Patched; later exploited as PHP2Shell PHP RCE flaws
PHP2Shell exploitation Attacker chained a custom exploit with a kernel flaw; AI tooling found in attack infrastructure, no direct evidence AI wrote the exploit Active intrusions reported; also chained with Check Point Gaia zero-day CVE-2026-123039 (CISA KEV Nov 13 2026), no AI role reported for that step PHP2Shell intrusions
Xen XSA-523 (CVE-2026-102703 / 102704) Found by Claude; affects every release since 4.6 Disclosed Nov 3 2026 Xen XSA-523
CVE-2026-114938 (Java4Shell, OpenJDK) Found with Codex in about an hour and exploited two hours later for under $200 in tokens (Calif’s claim) Patched Nov 17 2026; pre-auth RCE claimed in Calif tests; not known exploited in the wild (VulnCheck, Nov 17) Java4Shell
CVE-2026-111497 (Node2Shell) Found by OpenAI Codex Security researchers Reported Oct 23 2026 Node2Shell
CVE-2026-123158 (OpenSSH “sshole”) Found by a researcher using agent swarms; first exploit attempt under nine hours after disclosure. Root cause: 2016 OpenSSH fix plus musl behavior since 2012 (as reported) Patched Nov 5 2026; exploited; CISA KEV Nov 9 2026 OpenSSH sshole
CVE-2026-40217 (LiteLLM guardrail sandbox escape) Target is an AI tool: code execution in an LLM gateway; no report that AI found or exploited it Fixed in 1.83.10; exploited; CISA KEV Nov 18 2026 LiteLLM CVE-2026-40217
CVE-2026-90436 (Apache Axis 1.x, in Confluence and N-central) Rapid7 says the attacker’s command structure suggests AI-assisted tooling and later scripts were likely AI-generated End of life, no 1.x fix; exploited from Nov 13; CISA KEV Nov 18 2026 Apache Axis RCE
CVE-2026-102895 (GPUStack) Target is an AI tool: auth bypass in a GPU cluster manager for AI model serving; no report that AI found or exploited it Exploited; CISA KEV Nov 9 2026; fixed in 2.2.0 GPUStack auth bypass
CVE-2026-20320 (Cisco BroadWorks XSP XXE) Exploited by a jailbroken DeepSeek Flash agent in Operation Snowball (Malbear Labs); flaw patched months earlier Patched Aug 2026; CISA KEV Nov 10 2026 Operation Snowball
CVE-2026-124194 (Outlook NTLMv2 leak) One finder’s credit mentions use of Claude; role not stated Exploited zero-day; patched and CISA KEV Nov 10 2026 Outlook CVE-2026-124194
CVE-2026-81578 / 82078 (PaperCut NG/MF) Attacker used AI to build and test exploits, then hundreds of agents on Codex with a DeepSeek model hit 440 servers (GreyNoise) Zero-days, disclosed Aug 27 2026; mass-exploited Aug 31 PaperCut AI campaign
n-day web app flaws (CVE IDs not given) exploit_agent framework diffed patches and auto-ran AI-drafted exploits with no human in the loop (Google GTIG) Reported Nov 16 2026 exploit_agent
CVE-2026-102489 / 102490 (Zammad) Zero-days chained in a breach by an apparently autonomous agent Reported DIVD breach
CVE-2026-114713 (UnrealIRCd labeled-response) Found in a maintainer audit done with AI help; exploitation reported two days after the fix Fixed Nov 7 2026; exploited; CISA KEV Nov 13 2026 UnrealIRCd CVE-2026-114713
CVE-2025-3248 (Langflow) In the exploit playbook of [[artex ARTEX]] agents stealing AI provider keys (Amazon MadPot records; outcomes not verified by Amazon) Exploited by AI agents, Oct-Nov 2026
CVE-2026-28797 (RAGFlow) Same ARTEX playbook as above Exploited by AI agents, Oct-Nov 2026 ARTEX MadPot campaigns
66 TCP/IP stack flaws (Contiki, lwIP, wolfIP, Ethernut, proprietary; CVE IDs not given) Found by Forescout’s Claude-based harness for $1,200-$2,300 in tokens (vendor estimate) 54 reported; partly fixed Forescout TCP/IP bugs

Disclosure windows are shrinking

On Nov 16 2026 Anthropic cut its coordinated disclosure deadline from 90 to 30 days for high- and critical-severity bugs (45 for others) from Jan 1 2027, saying it sees more rediscovery of the same bugs as models improve and that a 90-day window assumed a flaw could stay private for months. The same day Google said attackers may find it more efficient to point LLMs at already-patched flaws than to hunt zero-days (exploit_agent framework).

Where the wiki looks

The daily update checks the CISA Known Exploited Vulnerabilities catalog for new entries and for any with reporting that AI found or exploited them, alongside vendor and researcher advisories.

Timeline

Every dated event in the wiki that involves AI and CVEs: latent flaws, now exploitable, newest first.

  • Policy
    Anthropic cuts its vulnerability disclosure window from 90 to 30 days, citing AI rediscovery

    For reports from Jan 1 2027, Anthropic's coordinated disclosure policy gives 30 days for high- and critical-severity issues and 45 days for others, down from 90; it will escalate unresponsive cases after 15 days and publish full details 30 days after a patch. The 7-day window for actively exploited bugs is unchanged. Anthropic says it is seeing more rediscovery of the same bugs as models improve, so a flaw can no longer be expected to stay private for months. See AI and CVEs: latent flaws, now exploitable.

Connections

AI and CVEs: latent flaws, now exploitable and everything one link away. Drag to move, scroll to zoom, click a node to open it.Open in the full graph →

What links here

Concept AI hacking event

Incident AI-discovered PHP RCE flaws (CVE-2026-109618/109619) · Apache Axis RCE (CVE-2026-90436) · ARTEX campaigns seen by Amazon MadPot (Oct-Nov 2026) · European Commission ARTEX attack · exploit_agent framework · Forescout AI harness finds 66 TCP/IP stack bugs · GPUStack authentication bypass (CVE-2026-102895) · Java4Shell (CVE-2026-114938) · OpenSSH "sshole" login bypass (CVE-2026-123158) · Operation Snowball · Outlook CVE-2026-124194 · PaperCut AI agent campaign · UnrealIRCd CVE-2026-114713

AI lab Anthropic

Company Calif · Rapid7

Sources

  1. AFP, “Websites go dark worldwide as AI exposes critical security flaws”, Oct 22 2026 (via The Korea Times)
  2. Help Net Security, “OpenSSH flaw exploited in the wild just hours after disclosure (CVE-2026-123158)”, Nov 6 2026
  3. CISA, “Known Exploited Vulnerabilities Catalog (JSON feed)”, Nov 10 2026
  4. BleepingComputer, “CISA flags Cisco BroadWorks flaw as actively exploited in attacks”, Nov 11 2026
  5. SecurityWeek, “Exploited Outlook Zero-Day Among 1,094 Flaws Patched by Microsoft”, Nov 10 2026
  6. oss-security, “UnrealIRCd 6.2.7.1 released & hot-patch (CVE-2026-114713)”, Nov 8 2026
  7. AWS Security Blog, “Amazon Threat Intelligence observes threat actors running cybercrime campaigns with ARTEX AI agents”, Nov 13 2026
  8. Forescout, “Finding Critical Infrastructure Bugs with Off-the-Shelf AI”, Nov 12 2026
  9. GreyNoise, “Agents Gone Wild: An AI-Orchestrated Global Campaign Against PaperCut NG/MF”, Sep 9 2026
  10. SiliconANGLE, “Google says an AI framework turned security patches into exploits with no human in the loop”, Nov 16 2026
  11. Anthropic, “Updating our coordinated vulnerability disclosure policy timelines”, Nov 16 2026
  12. BleepingComputer, “Check Point warns of zero-day flaw exploited for root access”, Nov 13 2026
  13. Calif, “Java4Shell: Write once, RCE everywhere”, Nov 17 2026
  14. Rapid7, “Rapid7 Detected Exploitation of CVE-2026-90436 against Atlassian Confluence Data Center, N-able N-central also Exploited”, Nov 16 2026
  15. GitHub, “Sandbox escape in custom-code guardrail (GHSA-wxxx-gvqv-xp7p)”, May 7 2026