Agentic AI weaponization inside breaches
Key Questions
What evidence shows AI is being used to discover zero-day vulnerabilities?
GTIG confirmed the first wild AI-assisted zero-day, while Sysdig documented the first fully autonomous AI intrusion. Anonymous researchers like 'bikini' and Exploitarium have publicly dumped AI-discovered zero-days affecting over 15 products including libssh2, Gitea, and FFmpeg.
What is Jadepuffer and how does it demonstrate agentic ransomware?
Jadepuffer is the first documented autonomous ransomware attack that used an AI agent to exploit Langflow and Nacos, steal credentials, and encrypt 1,342 config records while adapting in 31 seconds. It highlights the shift toward fully AI-driven attack chains without human intervention.
Which critical vulnerabilities were recently exploited shortly after disclosure?
Adobe ColdFusion CVE-2026-48282 was exploited within two days of patching, and Citrix NetScaler CVE-2026-8451 saw active exploitation within 48 hours. These cases underscore the rapid weaponization of high-severity flaws.
How effective are current LLMs at detecting vulnerabilities according to benchmarks?
A context-aware benchmark found the best-performing agent, Claude-3.7-Sonnet, achieved only 23.83% F1 score in vulnerability detection. This serves as a reality check against hype around tools like Mythos.
What new defense strategies are recommended against AI-discovered vulnerabilities?
Organizations are advised to shorten patch cycles, deploy AI defensively, automate incident response, and revisit disclosure policies. Apple has already shifted to rapid unbundled security updates in response to AI-driven discovery speeds.
What supply-chain risks were identified in AI-related open-source tools?
Critical unauthenticated RCE vulnerabilities in Langflow (CVE-2026-10134 and others) via Python injection were disclosed with validated PoCs, posing high risk to AI pipeline supply chains. No active exploitation has been observed yet.
Which older vulnerabilities were uncovered during recent AI-driven scans?
Project Glasswing identified 23,000 bugs including a 27-year-old flaw in OpenBSD. A human researcher also found the Bad Epoll Linux kernel LPE that Mythos missed, affecting Android with 99% success rate.
How is the NVD collapse affecting vulnerability prioritization?
Combined with the AI-driven wave of discoveries, the NVD issues have created a systemic prioritization crisis where critical vulnerabilities have doubled but only 7.8% require urgent action according to Check Point's report.
GTIG confirms first wild AI-assisted zero-day; Sysdig documents first fully autonomous AI intrusion; Project Glasswing finds 23k bugs including 27-year-old OpenBSD bug. Anonymous researcher 'bikini' dumps AI-discovered zero-days for 15+ products. Trump lifts export controls on Anthropic Fable 5 and Mythos 5 after new safeguards. Check Point Exposure Gap Report finds critical vulns doubled, only 7.8% need urgent action. Adobe patches 7 CVSS 10.0 flaws citing AI-driven discovery. AI vulnerability discovery now targeting banking infrastructure. NVD collapse combined with AI wave creates prioritization crisis. Apple AI-assisted patches for zero-days. Exploitarium researcher dumps 30+ AI-discovered zero-days (libssh2, Gitea, FFmpeg) without CVD; libssh2 CVE-2026-55200 already exploited. Apple shifts to rapid unbundled security updates due to AI-driven vulnerability discovery (Mythos). AI attack vectors article details Grok prompt injection, Galileu legal deception, under-8-minute AWS takeover via AI agent exploitation. Critical Langflow OSS vulnerabilities (CVE-2026-10134 etc.) with unauthenticated RCE via Python injection — PoCs validated, no active exploitation yet but high risk for AI pipeline supply chain. Jadepuffer autonomous ransomware attack — first documented agentic ransomware using AI agent to exploit Langflow and Nacos, steal creds, encrypt 1,342 config records, adapt in 31 seconds. Practical defense guide for AI-discovered vulnerabilities emphasizes shortening patch cycles, using AI defensively, automating IR, and revisiting disclosure policies. Academic study on automation bias in SOCs identifies two analyst profiles with distinct vulnerabilities — timely for SOC team composition amid rapid AI integration. Context-aware benchmark of LLMs for vulnerability detection shows best agent (Claude-3.7-Sonnet) only 23.83% F1, a reality check amid Mythos hype. Bad Epoll Linux kernel LPE (CVE-2026-???) discovered by human researcher, Mythos missed it — 99% success rate, Android impact, no active exploitation yet. Adobe ColdFusion CVE-2026-48282 exploited within 2 days of patch. Citrix NetScaler CVE-2026-8451 actively exploited within 48 hours. Weekly threat intel roundup confirms CISA KEV additions (SimpleHelp, SharePoint) and active zero-days (ColdFusion, Progress ADC, Langflow). New: FLARE-AI formal vulnerability disclosure mechanism for AI flaws, bridging coordination gap. New: Two-Minute Exploit article highlights AI collapsing skill floor for app attacks — 87% of apps attacked in 2026, client-side attacks invisible. New: OSERA and Akrites (Linux Foundation) as remediation infrastructure for AI-accelerated discovery. New: JadePuffer AI-driven ransomware confirmed in Vulnerability Intel Report. New: BeyondTrust RS/PRA critical auth bypass (CVE-2026-???, CVSS 9.2) found internally with AI assistance — no active exploitation yet but history of webshell attacks. New: Chinese APT targeting US/Canadian university physics departments via Roundcube XSS/deserialization chain, using AI-written IceCube stealer.