AI Security — week of 2026-07-06

JadePuffer: AI agent drives an end-to-end ransomware attack

BleepingComputer · 2026-07-04 · corroborated · reporting

JadePuffer is reported as the first ransomware to use an AI agent to automate the full attack chain end to end, raising evasiveness and speed.

Also: The Register Sec · The Hacker News

Cursor AI editor flaws

The Hacker News · 2026-07-01 · single-source · reporting

Cursor AI editor vulnerabilities allow prompt injection to escape sandbox, posing a significant security risk.

AI-Hallucinated Domains

The Hacker News · 2026-07-01 · single-source · reporting

Attackers utilize AI-generated domains for phishing, highlighting the need for enhanced domain validation.

Langflow RCE exploited in the wild

The Hacker News · 2026-06-30 · corroborated · reporting

The Langflow RCE flaw is being actively exploited against AI apps — to deploy Monero miners and as an AI-agent attack vector.

Also: The Hacker News

AI coding agents vulnerable

The Hacker News · 2026-06-30 · single-source · reporting

Decades-old shell trick bypasses safety checks in AI coding agents, demonstrating the need for robust security measures.

iOS AI Apps Leak API Keys

The Hacker News · 2026-06-30 · single-source · reporting

Exposed API keys in iOS AI apps allow unauthorized model requests, highlighting the importance of secure key management.

BioShocking

The Hacker News · 2026-06-30 · single-source · reporting

AI browsers leak credentials via gaming trick, demonstrating the need for enhanced browser security.

Newer models invent fields

Simon Willison · 2026-07-04 · single-source · research

Newer AI models can invent fields, potentially causing tool calls to fail and emphasizing the need for adaptability in AI development.

Poisoned MCP descriptions

The Hacker News · 2026-06-30 · single-source · reporting

AI agents can leak data without triggering alarms through poisoned MCP descriptions, highlighting the importance of robust monitoring.

Claude Fable 5 restored

The Hacker News · 2026-07-01 · single-source · reporting

Jailbreak-linked export controls lifted for Claude Fable 5, potentially impacting the security landscape.

AI assistant jailbroken

The Register Sec · 2026-07-01 · single-source · reporting

Trusted AI assistants can be turned against users, emphasizing the need for robust security measures in AI development.

LLMs vulnerable to prompt injection

The Register Sec · 2026-06-29 · single-source · reporting

Large language models are vulnerable to prompt injection, allowing attackers to trick them into giving sensitive information.