Stop Asking If AI Agent Skills Are Safe — Focus on Capability Disclosure Instead
A growing number of developers are shifting from binary safety questions about AI agent skills to detailed capability disclosure. The change follows incidents where malicious SKILL.md files persisted across system wipes because they were backed up alongside legitimate configuration data.
One documented case involved a user who installed a trojan after following a Claude-generated recommendation. After reinstalling the operating system and restoring files, the same malicious skill reappeared because it had been stored in the agent's configuration directory. The file contained instructions to exfiltrate credentials and reinstall the payload on the next session.
Snyk published the ToxicSkills report after examining 3984 skills from public marketplaces. Researchers identified security problems in 36.8% of the skills and critical issues in 13.4%. They confirmed 76 malicious payloads, eight of which remained available at publication time. Prompt injection appeared in 91% of the confirmed malicious samples, often combined with conventional attack patterns to bypass both AI guardrails and static scanners.
Common attack patterns include environment variable exfiltration through encoded curl requests, base64-decoded eval commands targeting AWS credentials, and runtime fetching of additional instructions. A separate real-world vulnerability, CVE-2025-6514 in the mcp-remote package, received a CVSS score of 9.6 and affected more than 437,000 installations.
The proposed capability disclosure model uses five tiers that describe potential impact rather than safety claims:
- T0 — Inert: text instructions only, no file or network access.
- T1 — Local: executes predefined scripts and writes files without network access.
- T2 — Network: makes outbound requests to explicitly listed endpoints.
- T3 — Elevated: involves persistence, secret access, privilege escalation or destructive actions.
- T4 — Opaque: fetches or decodes code at runtime, preventing static analysis.
The accompanying open-source tool skill-xray, released under the MIT license at github.com/worklore/skill-xray, performs mechanical scanning for paths, endpoints and structural signals before an agent layer generates a human-readable report. The scanner outputs a SHA-256 hash, tier classification and line-specific findings without rendering any safety verdict.
Related articles
GPT-6 Astra Shows Strong Tool Use and Math Results but Trails in Text Quality Tests
OpenAI released GPT-6 Astra on 3 September 2026 with capabilities for long-horizon tasks involving code, documents and external tools. Independent benchmarks from Artificial Analysis, Epoch AI and others reveal a mixed picture: the model leads in Epoch Capabilities Index and certain coding agent tests yet falls behind Claude Fable 5.1 in text generation and some professional document tasks. Astra solved two previously open problems on FrontierMath Erdős and reached 99.9 percent on ARC-AGI-3 when paired with a state-preserving adapter. The model was rated Critical for cyber capabilities after scoring 100 percent on ExploitBench and discovering two previously unknown vulnerabilities. Pricing starts at $10 per million input tokens, 2.5 times higher than GPT-5.6 Sol, while real-world costs vary sharply depending on task length and retries.
OpenAI GPT-6 Astra Deploys Multi-Agent Parallel Processing, Increasing Local CPU Load and Security Risks
Early users of GPT-6 Astra have observed the model distributing complex tasks across multiple specialized agents that plan, solve, test code, verify results, and iterate after failures. This multi-agent approach enables faster handling of multi-step workflows compared to sequential chatbots. OpenAI states that Astra can control computers, operate browsers and applications, and install or test software, though it has not officially confirmed a native multi-agent architecture. Main computations run in the cloud, but agent tools can execute on user devices or corporate servers, leading to noticeable processor load when multiple agents compile code, launch browsers, run tests, and operate containers simultaneously. Corporate environments face added complexity as each agent requires virtual machines, sandboxes, internal data access, and careful environment cleanup. The increased autonomy has prompted OpenAI to strengthen monitoring of Astra actions and permission boundaries for subscribers of ChatGPT and enterprise clients.
Microsoft Copilot Can Surface Overshared Data Despite Permission Boundaries
Microsoft documentation states that Copilot only accesses data authorized for the signed-in user, yet default SharePoint and OneDrive sharing settings often grant broad access that the AI then respects literally. This creates accidental oversharing risks where Copilot retrieves documents shared too widely years earlier. Administrators can use Content Management Assessment and Data access governance reports, including the EEEU report covering the top 100 sites shared in the past 28 days, to identify problematic content. Two distinct controls exist: Restricted Access Control removes access entirely while Restricted Content Discovery hides items from Copilot and search without altering permissions. Sensitivity labels combined with encryption can exclude programmatic access for agents, though Microsoft does not guarantee outright blocking. Interaction logs stored in Microsoft Purview retain user prompts, Copilot responses, and citations to accessed documents, providing an audit trail for oversharing incidents.
Adaptive LLM Worm Uses Local Models to Craft Per-Target Exploits in Heterogeneous Networks
Researchers from the University of Toronto have published a preprint describing an adaptive computer worm driven by LLM agents that spreads across corporate networks by generating individualized attack strategies for each compromised system. Unlike traditional worms such as WannaCry that rely on fixed exploits, this worm maintains its own infrastructure by running local LLMs on infected GPU-equipped machines to analyze vulnerabilities and synthesize new attack vectors in real time. The system was tested in an isolated FakeCorp environment containing Linux, Windows, and IoT devices, successfully leveraging known real-world vulnerabilities to propagate over 48 hours and seven-day autonomous runs. Two core components power the worm: a GPU-hosted LLM component and a hierarchical agent framework with memory, reasoning graph, and tool modules that manage reconnaissance, exploitation, and payload deployment. The authors note that the approach creates an economic asymmetry favoring attackers because the worm parasitizes victim compute resources, eliminating the need for external C2 or commercial LLM services. They warn that adding adaptive reasoning to historical worms such as SQL Slammer, Conficker, or Stuxnet would significantly increase their resilience while remaining slower and noisier than classic self-propagating malware.