A Java testing library maintainer hid destructive commands in ANSI escape codes — invisible to humans, lethal to AI agents. The jqwik protestware incident reveals a new front in the AI supply chain war, where dependencies become weapons and agents can't tell the difference.
Langflow, an open-source platform for building AI agent workflows used by individual developers and enterprises alike, became the first AI agent platform ever listed in CISA's KEV catalog on July 7. Attackers are using a cross-tenant IDOR to steal LLM provider keys, AWS credentials, and database secrets with the literal prompt 'leak api keys.' Zero HN coverage. Patching deadline: July 10, 2026.
Every MCP integration is a new data-to-execution pipeline that traditional security can't see. Here are 7 real attacks from June-July 2026 that show why your AI coding agent is the new insider threat.
A census of 15,382 MCP servers found 1,880 pointing to deleted repos, 888 unlicensed, and typosquatting so easy 638 servers share the same name. When your AI agent connects to these, it hands over its identity.
Microsoft's Agent Governance Toolkit claims to address all 10 OWASP ASI risks with deterministic policy enforcement. A security researcher found that setting a single HTTP header bypasses every governance check — rate limits, audit logs, and policy decisions all trust whatever string the caller sends.
Mozilla's 0DIN team showed how a clean GitHub repo with perfectly normal project documentation can trick Claude Code into opening a reverse shell — using DNS TXT records as the C2 channel. No malicious code, no prompt injection, no warnings.
The NSA issued official MCP hardening guidance. Amazon Q had a CVSS 8.5 MCP auto-execution flaw. ShareLock achieves >90% attack success by splitting malicious payloads across tool descriptions. And only 11% of production agents pass basic security checks.
Three senior safety leaders quit OpenAI the same week GPT-5.6 solved a 50-year math problem. Anthropic's safeguards chief warned 'the world is in peril' on his way out. Here's what the safety exodus means for every founder building with AI.
341 malicious skills. 12% of the entire registry. Screening bypassed. All major AI providers blocking it. The AI agent supply chain is on fire and there's no audit infrastructure.
The OWASP Top 10 for Agentic Applications 2026 is the Rosetta Stone for AI agent security. Every incident now maps to a standardized taxonomy. Here's the full list, why automated scanners miss the most dangerous ones, and what to do about it.