#ai-security

Researchers took over OpenAI staff ChatGPT accounts via a forum image bug and an SSO flaw, then reached internal repos via Codex. A checklist for connector use.

SentinelLABS matched public Hugging Face commits to OpenAI's May agent incident, to the second. What the report shows, and the logs a platform should keep.

GreyNoise traced an AI-agent campaign that hit at least 440 PaperCut servers in 48 countries, days after a patch existed. What it changes about patch order.

The NSA, CISA and FBI name six AI companies in advisory AA26-251A. See what changes for model procurement, provenance checks and API abuse monitoring.

OpenAI committed $1 billion in subsidised Daybreak access over six months. What it is, which organisations are prioritised, and what remains unpublished.

Google’s Fairwind joins Anthropic’s Glasswing and CVP, OpenAI’s Daybreak and Microsoft’s MDASH. One table of who is eligible for each cyber-capable model.

curl’s maintainer posted that Mythos and Codex Security had nothing left to find. Days later AISLE filed 29 reports; six became Low CVEs in curl 8.22.0.

Independent report: 1,200 OpenAI agents on a hidden message board, 700 joined the Hugging Face attack, 7% of reviewed transcripts were spoofed. What changes.

A census of 15 rows across 14 agent frameworks, scored against five SSRF controls from primary sources. Two default fetch paths re-validate every redirect hop.

A vendor-published disclosure ledger reports thousands of findings across hundreds of open-source projects. What maintainers and security teams should check.

A five-day plan to red-team AI agents with garak, promptfoo and PyRIT, mapped to the OWASP agentic categories and the MITRE ATLAS technique taxonomy.

Microsoft's MAI-Cyber-1-Flash claims 96% on CyberGym at half the cost. The routing thesis behind it matters more than the unverified benchmark.

Anthropic's free Claude Code security plugin scans code as it's written across three layers, and a new on-demand deep-scan plugin backs it with real patches.

Google's Gemini 3.5 Flash Cyber finds and patches vulnerabilities but stays gated. Why restricted access is becoming the default for dual-use AI models.

Hugging Face says an autonomous AI agent — not a human — ran an end-to-end intrusion of its infrastructure, stealing internal datasets and credentials.

Cybersecurity data for 2026: vulnerability exploitation overtook stolen credentials as the top breach entry point, and average breach cost fell to $4.44M.

Anthropic told US senators that Alibaba ran 28.8M Claude queries via 25,000 fake accounts, the largest model-distillation campaign it has disclosed.

OpenAI's Lockdown Mode severs the exfiltration stage of prompt injection, not injection entry. What it disables, who needs it, and how to deploy it RBAC-style.

Anthropic grew Project Glasswing to ~200 orgs in 15+ countries, with ICE/NYSE and Rubrik using its Mythos model to self-vet critical infrastructure.

A 12-layer defense framework for prompt injection — input sanitization, structured outputs, tool gating, output filtering, eval coverage, replay forensics.

Anthropic suffered two breaches in one week — Mythos document leak and Claude Code source leak. Enterprise AI security lessons and risk mitigation strategies.

OWASP's Agentic Top 10 defines the biggest security risks in autonomous AI systems. A plain-English guide covering each risk, real examples, and defenses.

1 in 8 enterprise security breaches now involve agentic AI systems. Threat landscape analysis with OWASP Agentic Top 10 mapping and defense strategies.

40% of business email compromise attacks are now AI-generated deepfakes. Detection strategies, employee training frameworks, and enterprise protection guide.

76% of organizations report unauthorized AI tool usage by employees. Shadow AI detection framework, governance policies, and risk mitigation strategies.

Anthropic accuses DeepSeek, Moonshot AI, and MiniMax of industrial-scale distillation via 24,000 fake accounts and 16M+ Claude exchanges. Full analysis inside.

The ClawHavoc attack exposed 341 malicious AI agent plugins. Essential security lessons for plugin ecosystems and agentic integrations.

OpenClaw's full system access creates significant attack surface. Complete security hardening guide with prompt injection defense and containerization.

341 malicious skills found on ClawHub in the ClawHavoc campaign. Full analysis of the attack, affected users, and VirusTotal partnership response.

Secure AI coding assistants in enterprise: data protection, code leakage prevention, compliance, and governance for Copilot, Claude, and Cursor.

Secure AI agents with enterprise best practices. 24 CVEs across top tools. Prompt injection, data exfiltration prevention. Complete guide.