Your Agent's Dependencies Are the New Attack Surface

A poisoned agent-framework release stole credentials on every Python start. Why agents need to re-check packages and skills.

PUBLISHED READ 6 MIN
A robot beside a conveyor and stacked parcels, with a snake-like wire rising from an opened box toward a ring of keys

An agent can reach whatever its code can reach. That code is mostly not yours: Python packages, npm modules, skills, plugins and MCP servers, pulled in at install and loaded on every run. This week showed what that means in practice. Two releases of a real Microsoft agent package shipped a credential stealer, and it ran before anyone imported a thing.

Last week we wrote that agent skills are the new npm packages and that scanning a skill once is not enough. This post widens the lens to everything an agent loads, and what changed in the last seven days.

What happened to Microsoft's agent-framework

On November 17, versions 1.26.1 and 1.26.2 of Microsoft's agent-framework package on PyPI were compromised with a credential stealer, according to Socket. This was not a look-alike name. It was the real package, under its real name, with new version numbers.

Three details matter:

  • It ran on every Python start. Each version shipped a .pth file. Python executes those when the interpreter starts, so the stealer ran whenever Python ran, including pip itself. Your code never had to import the package.
  • It went after exactly what agents hold. Environment variables containing TOKEN, SECRET, KEY, AWS, AZURE, GITHUB, OPENAI or HF_, other processes' environments, Azure managed identity tokens, GitHub Actions OIDC tokens, and config files for pip, npm, Git, Docker and Hugging Face, plus .env files.
  • The second version covered its tracks. In 1.26.2, Socket notes, "the legitimate package initializer is restored byte-for-byte." Anyone checking only the package's own code would see nothing wrong.

Both versions are now quarantined, and 1.26.0 is the latest clean release. If you installed either bad version anywhere, upgrading is not the fix. Rotate every credential that machine or CI runner could see.

Why agent dependencies are a better target

A normal library sits in an app with a database password. An agent framework sits in a process holding model API keys, a GitHub token, cloud credentials and often a wallet. It also runs in CI, where OIDC tokens can be traded for cloud access. One poisoned release in that position is worth more than a hundred in a web app.

And agents load more than packages. A 2026 academic study of 31,132 agent skills, cited by NVIDIA's SkillSpector scanner, found that 26.1% contained at least one vulnerability and 5.2% showed patterns that strongly suggest malicious intent. Skills, plugins and MCP servers are dependencies too. They just skip the package manager.

Three signals from this week

The trusted name was the attack. Every check that asks "is this the package I meant?" passed. The name was right. Only the version was new, and the payload hid in a file type most reviews never open.

GitHub now treats agent setup as supply chain. On November 16 GitHub added a Copilot setting that lets enterprises block skills, custom agents, hooks and MCP servers that do not come through plugins. GitHub is careful to say it "isn't a marketplace allowlist" and should be paired with its approved marketplace controls. The direction is clear anyway: where an agent's capabilities come from is now a policy question, like where packages come from.

AI scanning produces noise unless someone proves it. GitHub and Microsoft ran AI scanning across projects like Express, Caddy, FastMCP and AutoGPT. It flagged 9,791 possible issues. After triage and manual review, 99 became private reports, and maintainers had fixed 55 of them by November 15. About one raw finding in a hundred was worth a maintainer's time. Their title says it: maintainers don't need more AI bug reports, they need fixes.

Why one scan at install is not enough

A scan at install answers a narrow question: was this version clean, by these rules, today? This week broke all three parts of that.

  • The version changes. agent-framework was trustworthy at 1.26.0. A lockfile without hashes, or a range like >=1.26, picks up 1.26.1 on the next build.
  • The payload hides from the rules. A .pth file, an initializer restored to its original bytes, or invisible characters in a skill all pass a check that looks in the usual places.
  • The rules change. When a new trick is published, everything you installed before was checked without it. Nothing re-reads it unless you schedule it.

So the question is not "was it scanned?" It is "when was it last checked, against which rules, and what changed since?"

What Aeon does today

Aeon is an agent that runs on GitHub Actions, so it has the same exposure. Here is what is in place, and nothing more.

Skills are checked at install, including for hidden characters. scripts/skill-scan.sh gates every skill install, and since November 12 it flags Unicode TAG smuggling, bidi overrides and zero-width characters, the tricks that make an instruction invisible to a human reviewer and visible to the model.

Scans run on a schedule, not once. Aeon's vulnerability scanner can run every week on a cron, and Aeon scans its own codebase the same way. A rule added today applies to code installed last month.

Findings are proven before anyone sees them. That is the gap the GitHub numbers expose. Aeon's scanner tries to refute each candidate and keeps only what it can reproduce. The result so far: 153 vulnerabilities found and fixed in 113 open-source projects.

Agent tooling is part of the record. Our VulnScanning series covered real cases, each reported privately and patched by its maintainers. In Episode 5, a Kaneo integration returned its Gitea token to any workspace member, and ds2api wrote API keys into its logs. In Episode 6, the shell parser behind Omnigent's agent guardrails failed open (CVE-2026-62676), and opencodex had a stream buffer with no size limit. Omnigent, opencodex and ds2api are tools agents run on, which makes them agent dependencies too.

What to do with your own agent

  • Check for agent-framework 1.26.1 or 1.26.2. If either was installed, rotate every secret that machine or runner could read, including CI OIDC trust.
  • Pin with hashes. Lock exact versions and hashes so a new release cannot slide in on the next build.
  • Wait before adopting new releases. Many poisoned versions are caught within hours. A delay of a few days on new versions costs little.
  • Keep secrets out of install steps. Dependency installs and builds should not see your model keys or deploy tokens.
  • Re-scan on a schedule. Packages, skills, plugins and MCP servers, every time your rules change and at least weekly.
  • Control where capabilities come from. If your platform has an approved source setting for skills, plugins or MCP servers, turn it on.

FAQ

What happened to Microsoft's agent-framework package?

On November 17, 2026, versions 1.26.1 and 1.26.2 on PyPI were compromised with a credential stealer that ran on every Python start through a .pth file. Both are quarantined. 1.26.0 is the latest clean release.

What is a .pth file and why is it dangerous?

A file in Python's site-packages folder that Python reads at startup. Lines starting with import are executed, so a malicious package can run code without ever being imported.

Is a dependency or skill scanner enough?

No. A scan checks one version against today's rules. Pair it with hash pinning, a delay on new releases, scoped secrets and scheduled re-scans.

Try it

Aeon is open source and MIT licensed. Fork it at github.com/aeonfun/aeon, or set it up in the browser with Aeon Connect. Every fork ships with the skill install gate, and the vulnerability scanner is one skill away.