High-signal AI/security/automation notes.
Mohan Pedhapati of Hacktron used Claude Opus 4.6 to produce a working Chrome V8 exploit chain for $2,283 in API costs, demonstrating that widely available AI models can weaponize vulnerabilities.
A CSA/SANS/OWASP collaborative paper with 250 CISO contributors outlines a Mythos-ready security program to counter AI-driven vulnerability discovery.
LMDeploy CVE-2026-33626 is a server-side request forgery vulnerability in the vision-language module that fetches arbitrary URLs without IP validation, risking cloud metadata exposure.
New arxiv paper demonstrates that LLM debugging agents consuming cloud logs are vulnerable to indirect prompt injection embedded in log entries — with cloud vendor guardrails failing to detect the attacks.
AI social media scheduling tool Postiz ships a stored XSS vulnerability (CVSS 8.9) allowing authenticated users to upload executable content via Content-Type header spoofing — fixed in v2.21.6.
The UK AI Safety Institute found all tested frontier LLMs remain highly vulnerable to basic jailbreaks, with several completing high-school-level cyber challenges.
Palo Alto Networks Unit 42 warns that frontier AI models can now function as full-spectrum security researchers, autonomously discovering zero-days and chaining exploit paths.
Dark web intelligence reveals underground operators are using Claude, Gemini and ChatGPT with better prompts — not ablated models like WormGPT — for exploit development and payload generation.
Vercel confirmed a security breach after a compromised Google Workspace OAuth app belonging to a third-party AI tool (Context.ai) gave attackers access to internal systems and customer environment variables.
Georgia Tech researchers track 74 confirmed CVEs directly caused by AI-assisted coding tools, with 35 disclosed in March 2026 alone — and estimate the real number is 5–10× higher.
arXiv:2604.11790 introduces ClawGuard, a runtime security framework that enforces deterministic rule-based controls at every tool-call boundary to block indirect prompt injection across web, MCP, and skill-file attack vectors.
GreyNoise honeypots captured 91,403 attack sessions targeting LLM infrastructure, including systematic enumeration of 73+ model endpoints and SSRF campaigns exploiting Ollama.
iProov Threat Intelligence Report 2026 documents a 1,151% surge in iOS-targeted injection attacks using fake video and biometric data to bypass identity verification systems, alongside a 720% spike in Southeast Asia.
arXiv:2603.28013 — MIT and University of Chicago researchers track prompt injection through four kill-chain stages across five frontier models, revealing that write-node placement is the highest-leverage safety decision in agent pipelines.