Posts
High-signal AI/security/automation notes.
ExploitBench — AI Agents Achieve Arbitrary Code Execution on V8
Carnegie Mellon researchers publish ExploitBench, showing Claude Mythos and GPT-5.5 can autonomously build browser exploits from known V8 bugs.
Linus Torvalds — AI Bug Reports Overwhelm Linux Security Mailing List
Linus Torvalds says AI-generated bug reports are flooding the Linux security mailing list, prompting new triage rules for automated vulnerability discovery.
MCPSafe — 7 Coordinated Disclosures After Scanning 50+ MCP Servers
MCPSafe scanned 50+ MCP server repositories and found indirect prompt injection, SSRF, and annotation mislabeling across official servers from Anthropic, GitHub, Atlassian, Microsoft, and more.
OpenClaw Five-Point Security Plan — fs-safe, Proxyline, ClawHub Ratings
OpenClaw unveils a five-point security plan including fs-safe filesystem protection, Proxyline network proxy, ClawHub trust ratings, smarter confirmation dialogs, and automated OpenGrep checks.
Pwn2Own Berlin 2026 — OpenAI Codex Exploited, $1.29M in 47 Zero-Days
Pwn2Own Berlin 2026 concluded with $1.29M awarded for 47 zero-days, including a successful exploit against OpenAI Codex AI coding assistant.
arXiv — MATRA Threat Modeling Framework for Agentic AI Systems
MITRE researchers present MATRA, a pragmatic threat modeling framework that adapts established risk assessment methodology to quantify agentic AI deployment risks using OpenClaw as a case study.
arXiv — Securing AI Agents Like Operating Systems
Researchers argue that LLM-based agents face the same isolation, privilege separation, and communication mediation challenges as operating systems, and demonstrate that many protection mechanisms fail in practice.
arXiv — Security Risks in Tool-Enabled AI Agents in Privileged Cloud Environments
IEEE COMPSAC 2026 short paper analyzes security risks of cloud-hosted AI agents, finding that most risks arise not from novel vulnerabilities but from over-privileged tools, capability-intent mismatches, and ambient authority leakage.
Cymulate — Prompt Injection Triggers Zero-Click RCE in AI CLI Tools (Cursor, Kiro, Codex, Gemini)
Cymulate researchers uncovered zero-click RCE chains across Cursor CLI, AWS Kiro, Codex Desktop, and Gemini CLI — where a single prompt injection can silently execute code via Windows PATH hijacking and config file poisoning.
EU CRA — Vulnerability Reporting Obligations for AI Vendors Begin September 2026
Greg KH and the oss-sec mailing list clarify that EU Cyber Resilience Act Article 14 vulnerability reporting obligations for AI manufacturers take effect September 11, 2026.
Forcepoint — 10 In-the-Wild Indirect Prompt Injection Payloads Targeting AI Agents
Forcepoint researchers found 10 real indirect prompt injection payloads hidden on live websites that instruct AI agents to delete files, steal API keys, and process fraudulent payments.
MCP Database Flaws — Apache Doris, Pinot and Alibaba RDS Vulnerabilities
Security researcher Tomer Peled found three critical MCP server vulnerabilities affecting Apache Doris, Pinot, and Alibaba RDS — including SQL injection and metadata exfiltration — with one vendor declining to patch.
Hunt.io — TeamPCP FIRESCALE Malware Uses GitHub Dead-Drop for C2 Resilience
A 13-file Python toolkit deployed by TeamPCP after npm/PyPI supply chain compromises uses a 3-tier exfiltration chain, a GitHub commit-based dead-drop (FIRESCALE), and targets AWS GovCloud credentials.
VectorSmuggle — Steganographic Data Exfiltration Through AI Embeddings
New VectorSmuggle framework demonstrates how attackers can hide arbitrary data inside vector embeddings using steganography, bypassing DLP tools and vector database security checks.
arXiv — Re-Triggering Safeguards: Embedding Disruption for Jailbreak Detection
New ICML 2026 paper proposes injecting noise into token embeddings to re-activate LLM safety safeguards, effectively detecting jailbreak prompts by testing their inherent fragility.
Cyera — Four Chainable OpenClaw CVEs Expose 180K+ AI Agent Servers
Cyera disclosed four chainable vulnerabilities in OpenClaw (CVSS 7.7–9.6), including a critical sandbox escape and privilege escalation, affecting 245K publicly exposed AI agent deployments.
Google GTIG — PROMPTSPY Autonomous AI Malware Interprets Systems and Generates Commands
Google Threat Intelligence Group reveals PROMPTSPY malware that uses embedded LLMs to interpret system states and dynamically generate commands, signaling a shift toward autonomous attack orchestration.
HackerOne — Prompt Injection Reports Surge 540% Year-over-Year
mlflow — CVE-2026-2652 Authentication Bypass Exposes ML Experiment Jobs
CVSS 8.6 auth bypass in mlflow ≤3.9.0 allows unauthenticated access to job submission and OpenTelemetry trace injection when served via uvicorn with basic-auth enabled.