Your AI agent's plugins are part of your supply chain
AI coding agents have quietly grown an ecosystem. Plugins, extensions, hooks and "skills" – small instruction packages an agent loads to learn a library or a team's way of working – can now be installed with a single command, and increasingly arrive with ordinary package updates. That is useful. It also means the code and text around your model has become part of your software supply chain, with the same permissions as the developer running the agent.
Three recent publications make the case from different angles.
What the sources show
CERT-EU tested backdoors in agent harnesses. On 30 October, CERT-EU's AI team described an experiment in an isolated lab. They added a dormant payload to local copies of two popular third-party packages for the Pi and Claude Code agent harnesses. The trigger was an ordinary review phrase, "lgtm, ship it". When it appeared, the extension collected host details, environment variables and credential-related files from the working directory. The model never had to produce a malicious response, and controls that restricted the agent's own file-reading tools did not cover what the plugin read directly. CERT-EU's conclusion: "controls that restrict the model's actions do not necessarily restrict third-party code" running alongside it. This was a controlled demonstration, not a reported attack, but the mechanism is the familiar one from IDE plugins and npm packages.
Skills are moving into normal package flows. On 3 November, the .NET team showed how NuGet packages can ship agent skills next to their binaries, so the skill matches the exact package version. Because dotnet restore already downloads packages, the skills land on every developer's machine; a separate tool can copy them into the repository for agents to use. The post is candid about the downside: agents turn markdown in packages into active content, "a new threat vector that needs to be considered". It also notes that poorly written skills can make results worse and more expensive.
A significant share of public skills has weak spots. Researchers at Politecnico di Torino built SkillSecurer, an LLM-based scanner for prompt-injection risks in skills, and ran it on 954 popular skills from the skills.sh catalogue. It flagged 168 of them (17.6%) with latent vulnerabilities. In a manual review of 100 findings, a majority of three reviewers confirmed 84. These are flagged weaknesses, not proven malware – but when the authors tested five flagged skills in a coding agent, 13 of 30 model-skill runs triggered the problem.
A fourth paper adds a detail worth remembering: in a study of where prompt injections land, an agent followed the same unsafe instruction about 2.6 times more often when it was placed in a README or TODO file than in a code or data file. Text that looks like guidance is trusted more.
Why this matters in regulated environments
For organisations under NIS2, none of this is a new category of obligation. Article 21 already lists "supply chain security" and "security in network and information systems acquisition, development and maintenance" among the required risk-management measures. Agent extensions simply belong in that scope – and in many organisations they are not in the inventory yet.
In healthcare and the public sector the stakes are concrete. A developer workstation or build agent often has credentials to test environments, package feeds and source control. A plugin that reads environment variables reads all of that, whether or not the model is well behaved.
What we would put in place
Inventory and pin. Treat agent plugins, hooks, MCP servers and skills as dependencies. Keep a list per team, pin versions, and review the diff before updating – including skills that arrive inside packages.
Make skills reviewable code. Skills that steer agents in your repositories should live in the repository, have an owner, and go through the same pull-request review as the code they influence. A skill that nobody can explain should not be loaded.
Isolate the agent, not just the model. Run agents in containers or dedicated VMs without access to host credential stores, SSH agents or browser profiles. Use short-lived, narrowly scoped credentials and an egress allow-list. This contains a bad extension even when nobody spots it.
Watch beyond tool calls. Approval prompts only see what the model asks for. Monitor extension installation, access to sensitive files and unexpected outbound connections at the workstation or container level.
Keep production data out of the blast radius. Agents that work on code should not run where patient data or production secrets are reachable. That separation is good practice anyway; agents make it urgent.
The trade-off
None of this means banning extensions. Skills can make agents more accurate and cheaper, and package-delivered skills that match the installed version are a sensible idea. The cost is governance: an internal approval flow, a curated feed or repository, and some friction when developers want the newest plugin today. For most regulated organisations that friction is cheaper than explaining afterwards why a review phrase in a pull request sent credentials somewhere.
The principle from CERT-EU is the one to keep: "Securing AI agents means securing their entire execution environment, not just the model itself."
Sources
- CERT-EU, "Beyond the model: backdoors in AI agent harnesses" (30 October 2026): https://cert.europa.eu/blog/beyond-the-model-backdoors-in-ai-agent-harnesses
- Richard Lander, .NET Blog, "Delivering Skills in NuGet packages" (3 November 2026): https://devblogs.microsoft.com/dotnet/delivering-skills-in-nuget-packages/
- Mecca et al., Politecnico di Torino, "SkillSecurer: Detecting and Patching Prompt-Injection Vulnerabilities in AI Agent Skills", arXiv:2609.14079 (September 2026): https://arxiv.org/abs/2609.14079
- Guo et al., "Where You Inject Matters: The Trust Region Effect in LLM Agents", arXiv:2610.38867 (October 2026): https://arxiv.org/abs/2610.38867
- Directive (EU) 2022/2555 (NIS2), Article 21: https://eur-lex.europa.eu/eli/dir/2022/2555/oj/eng