Safety, Alignment & InterpretabilityCisco (Amy Chang, Vineeth Sai Narajala, Idan Habler)

Cisco finds 26% of 31,000 agent skills have at least one vulnerability

An analysis of 31,000 agent skills found that 26% had at least one vulnerability. Documented silent data exfiltration, prompt injection, command injection and tool poisoning as types, and released an open-source Skill Scanner

Why it matters

The first large-scale measurement to treat skill marketplaces as an unvetted supply chain

Source — Primary sources first; secondary reports are labelled “Press”

Officialblogs.cisco.comhttps://blogs.cisco.com/ai/personal-ai-agents-like-openclaw-are-a-security-nightmare

Added in Week 35, 2026 · See that week

Same week

OpenClaw ships a fix for CVE-2026-25253PressOpenClaw name settled (Warelay→Clawdbot→Moltbot→OpenClaw)PressMoltbot is renamed OpenClaw and launches an official blogOfficialMoonshot AI releases Kimi K2.5Official

Nearby on the same axis

OpenClaw ships a fix for CVE-2026-25253PressOpenClaw name settled (Warelay→Clawdbot→Moltbot→OpenClaw)PressAttorneys general of 35 US states and territories send joint letter to xAIGovernmentAnthropic releases a fully revised Claude's ConstitutionOfficial