Archives

Agents working together to circumvent defenses raises critical questions about design versus emergent behavior

A MemTensor compromise weaponized GitHub Actions to steal publishing tokens and spread as a self-replicating worm via stolen credentials.

A campaign is poisoning the data that AI chatbots draw from, serving real users with scam contact details and phishing pages.

A new Windows malware outsources C2 decisions to multiple LLMs for autonomous deployment

A critical RCE in Bifrost exposes every connected LLM provider's credentials.

An undocumented setting in Meta's Muse lets local attackers intercept voice data without privilege escalation.

A financial services firm's accounting agent burned through $50,000 in cloud costs due to a corrupted data value with no cost guardrails.

SHA pinning fails when AI agents skip verification, leaving code supply chains exposed.

Two vulnerabilities in OpenAI's Codex reveal why software-level sandboxing fails against attackers with internal access.

Researchers used Claude Opus 5 to develop a working exploit in 72 hours, demonstrating real-world AI-assisted exploitation.

A threat actor uses LLM-written malware to harvest CI/CD secrets, then reports the vulnerabilities for rewards instead of selling the access.

Sandbox escapes in ChatGPT, DeepSeek Harness and Docker; a record Patch Tuesday; and OpenAI publishing six of its own misalignment incidents.

Critical symlink vulnerability in Docker Sandboxes allows guest code to access host files.

A zero-click RCE in top AI coding agents exploits a fundamental flaw in how they verify code commits.

Spain's data protection authority logged its first data breach executed by an autonomous AI agent, though the evidence remains limited.

A single browser extension can hijack AI assistants across Chrome, Edge, Opera, and Perplexity by exploiting a shared architectural flaw.

OpenAI published six internal misalignment incident reports and a disclosure framework on 16 September 2026. Most of the six describe missing egress controls and credential hygiene rather than novel model deception.

Hijacked AI assistant sessions become a deployment vector for supply chain worms, not just a data access risk.

WordPress deployed automated security gates with no human review fallback for plugin releases.

AI models are finding vulnerabilities faster than security teams can validate and patch them, creating a structural validation crisis.