Fetching from the wire…
Public story · 2026-02-25 · source-backed
(intermediate) — Audit for eval/exec/subprocess shell=True, replace with safe alternatives (ast.literal_eval, JSON parsers, execFile + arrays), implement allowlist validation, enable auth. (DEV Community / Kai Security)
Each link below shares sources, entities, or timing with this story.
The MCP CVE count reached 30+, all sharing the same root cause: user-controlled input reaching exec()/eval() without sanitization. AgentAudit scanned 194 MCP packages and found 118 security findings across 68 packages — 14 rated critical or high. Anthropic's own Git MCP server...
30 MCP CVEs Mapped Into Three Attack Layers — Comprehensive mapping: Layer 1 (Execution, 43%): 13 exec/shell injection CVEs. Layer 2 (Tooling, 20%): 6 CVEs targeting dev infrastructure — MCP Watch (a security scanner!) has command injection in its own repo cloning. Layer 3 (Ne...
Stop parsing JSON with regex. XGrammar modifies the probability distribution at every decoding step to force schema-valid output, compiling EBNF grammars to finite-state machines with near-zero overhead. OpenAI's structured output hits 100% schema validity; Anthropic reaches 9...
ECP captures agent outputs, tool invocations, and audit context uniformly, with adapters for LangChain, LlamaIndex, CrewAI, and PydanticAI so the same checks run against any of them. arXiv The authors explicitly label it work-in-progress with the method set expected to change....
Isolate the agent from the evaluator, never pass reference answers into the agent's context, never eval() untrusted model output, sanitize LLM-judge inputs against injection, and adversarially test the evaluator before trusting its scores (AI Agent Insights). A leaking or inje...
Domain: agent-security | Difficulty: advanced | Source Audit all MCP servers with uvx mcp-scan@latest. Implement network egress allowlists (ARXON succeeded via unrestricted outbound access). Scope tool permissions with PreToolUse hooks. Isolate knowledge base per-session. Moni...
MindPattern daily
One email a day at 7 AM. Sources and a take on every story. Unsubscribe anytime.