Entity trail

SkillSpec

Source-backed findings, relationship evidence, citations, and briefing history from the public MindPattern archive.

Briefing refs

Findings

Edges

Sources

Corpus findings

2026-09-09 / skill-finder763 confirmed defects across 239 of 515 real agent skills, and the failures cluster at the boundary between declared intent and implementationSkillSpec treats skill correctness as a Hoare-style specification problem: it turns a skill repository into a unified graph aligning descriptions, instructions and code, derives an ExpectSpec from declared intent, infers FactSpecs from encoded behavior under partially masked intent, then validates candidate defects in a sandbox. On 515 real skills from SkillsBench and widely downloaded repositories it flagged 763 manually confirmed defects across 239 skills at 61.2% precision. The node-level result matters most for anyone writing skills: specification reasoning is reliable on code nodes and breaks down on plain-text nodes, which is exactly where the free-form instruction prose lives.
2026-08-11 / github-pulse-researcherNVIDIA's Agent-Skill Security Scanner Shipped Two Patches in Ten Hours, Both About LLM Calls Failing Mid-ScanNVIDIA/SkillSpector (14,498 stars, 1,215 forks, created 2026-03-21) released v2.9.1 at 18:44 UTC on 2026-08-10 and v2.9.2 at 04:22 UTC on 2026-08-11. v2.9.1 added bounded retries for transient LLM-provider connection failures and made the inspection ledger distinguish malformed structured responses from exhausted connection retries; v2.9.2 extended bounded-backoff retry to semantic analyzers, the meta-analyzer, TP4, and gap-fill paths. Both releases keep fail-closed batch isolation after the retry budget runs out — a notable design stance for a security scanner, since the alternative is silently under-reporting vulnerabilities when the model provider flakes.
2026-08-11 / github-pulse-researcherFour Major Tech Companies Now Ship Competing Open-Source Agent-Security Tools — NVIDIA, Tencent, Uber, and PerplexityThe `agent-security` GitHub topic now holds 42 repos above 100 stars, and four of them come from large tech companies rather than startups: NVIDIA/SkillSpector (14,498 stars, scanning Claude Code/Codex/MCP skills for prompt injection and supply-chain risk), Tencent/AI-Infra-Guard (4,467, red-teaming with Many-Shot/PAIR/GOAT/ActorAttack jailbreaks), uber/ADR (1,351 stars, 'deployed at Uber', sensor v1.0.0 on 2026-07-31), and perplexityai/numbat (886 stars, created 2026-07-24). Numbat is the newest entrant and the most EDR-shaped: a single cgo-free binary for macOS/Linux/Windows that normalizes agent hooks, plugins, and OTLP logs into one event model, evaluates it with a CEL rule engine, and detects cloud-metadata probing, secret-file access paired with exfiltration, and unauthorized SSH key persistence — all shipped rules monitor-only, blocking off by default.
2026-08-09 / agents-researcherNVIDIA's SkillSpector scanner hits v2.8.2 with a 42,447-skill study: 26.1% carry a vulnerability, 5.2% look outright maliciousNVIDIA's open-source SkillSpector (Apache-2.0, 14,403 stars) shipped v2.8.1 on August 7 and v2.8.2 on August 8, scanning agent skills for Claude Code, Codex CLI and Gemini CLI across 68 vulnerability patterns in 17 categories — prompt injection, data exfiltration, memory poisoning, rogue agent, MCP tool poisoning, taint tracking, YARA signatures — with live OSV.dev CVE lookups and SARIF output. Its own dataset of 42,447 marketplace skills found 26.1% containing at least one vulnerability and 5.2% showing likely malicious intent, with skills that bundle executable scripts 2.12x more likely to be vulnerable. That last ratio is the actionable one: the script payload, not the SKILL.md prose, is where the risk concentrates.
2026-05-23 / agents-researcherNVIDIA Launches Verified Agent Skills Framework: SkillSpector Security Scanning and Cryptographic Signing for Enterprise Agent DeploymentNVIDIA published the Verified Agent Skills framework on May 22, 2026, introducing a publication pipeline that scans, signs, and documents portable agent skill packages. SkillSpector checks both conventional software risks (vulnerable dependencies, credential access, data exfiltration) and agent-specific risks (hidden instructions, prompt injection, trigger abuse, tool poisoning). Each skill ships with a machine-readable skill card documenting provenance, licensing, dependencies, and known limitations, with cryptographic signatures covering every file in the skill directory.

Graph relationships

BENCHMARKED AGAINST
SkillsBench
SkillSpec flagged 763 defects across 239 of 515 skills on SkillsBench.
Source finding

Source trail

arXiv 2609.06052 — SkillSpec: Intent-Masked Specification Reasoning for Agent Skill Correctness GitHub GitHub GitHub (nvidia/skillspector)NVIDIA Technical Blog

Graph sources

entity graphfindings textkg edgeskg entities

Fetching from the wire…