Weekly · News

AI Security — week of 2026-07-20

July 20, 2026 · 12 developments

Simon Willison · 2026-07-16 · corroborated · research

GPT-5.6 Codex has a critical bug that can delete files when unsandboxed.

Simon Willison · 2026-07-15 · corroborated · research

xAI’s Grok Build has been found to leak user directories and code repositories.

Simon Willison · 2026-07-15 · single-source · research

Claude LLM has a vulnerability that allows data exfiltration via web_fetch loophole.

The Hacker News · 2026-07-17 · single-source · reporting

A new botnet, NadMesh, has been discovered targeting exposed AI services for cloud keys.

The Hacker News · 2026-07-20 · single-source · reporting

Hugging Face has been breached by an autonomous AI agent.

The Hacker News · 2026-07-15 · single-source · reporting

The evolution of TuxBot v3, an LLM-assisted IoT botnet, poses new threats.

The Hacker News · 2026-07-16 · single-source · reporting

A new technique, Agent Data Injection, can corrupt AI agent inputs to run attacker commands.

The Hacker News · 2026-07-16 · single-source · reporting

A flaw in n8n’s token exchange allows attackers to log in as users from another issuer.

The Register Sec · 2026-07-16 · single-source · reporting

An open-weight AI model has been poisoned for under $100.

The Register Sec · 2026-07-14 · single-source · reporting

The Gemini AI model can be hijacked for malicious use.

The Register Sec · 2026-07-17 · single-source · reporting

Decades-old email tricks can still evade LLM email filters.

Simon Willison · 2026-07-17 · single-source · research

Kimi K3 has been found to refuse to leak system prompts.