ZYZ

AgentSec is all you need to make agent security legible. A2S (AgentSec Stack) breaks it into 6 layers: Identity, Cognition, Orchestration, Action,Enforcement, Evidence. Use it to design, evaluate, and govern AI agents without blind spots. Read the article: agentsec.vercel.app/blogs-in… Contribute to the community map: form.typeform.com/to/IzmSaNm… #AgentSec #A2S #AISecurity #AIAgents #LLMSecurity

🦞 1/ OpenClaw hit ~209K GitHub stars (as of Feb 19, 2026). Within weeks, researchers reported: • 40,214 internet‑exposed instances • ~1.5M tokens exposed via MoltBook • ~900 malicious skills flagged • 1‑click RCE: CVE‑2026‑25253 We did a full security teardown. Here’s what we found 👇 2/ We used A2S (AgentSec Stack) — 6 layers to assess agent security: L1 Identity · L2 Cognition · L3 Orchestration · L4 Action · L5 Enforcement · L6 Evidence 3/ OpenClaw shows gaps in almost every layer. But one is catastrophic: L5 — Enforcement. 4/ L5 is the non‑bypassable, fail‑closed gate between what the model decides and what the system does. Without it, “agent safety” becomes wishful thinking. 5/ The results (Wiz / SecurityScorecard / Bitdefender / NVD): • MoltBook token exposure via broken Supabase RLS • 40K+ exposed deployments in the wild • ~900 malicious skills flagged across the ecosystem • CVE‑2026‑25253: 1 click → arbitrary command execution 6/ Full deep dive (sources + every layer + mitigations): agentsec.vercel.app/blogs-in… #AgentSec #AISecurity #A2S #OpenClaw #MoltBook #AIAgents #PromptInjection #CVE #MCP