AI tools hold the keys now. A patch doesn't take them back.
We read CISA's KEV catalog (version 2026.11.13) and the advisories behind three pipeline compromises. Keys leaked within hours, federal due dates came 21 to 30 days later, and no KEV entry yet tells agencies to rotate.
- CISA's KEV catalog added 11 exploited flaws in AI model and agent tools in 2026, against 1 in 2025. Several expose the API keys and cloud secrets those tools hold.
- In three pipeline compromises, credentials leaked in hours. KEV due dates landed 21 to 30 days after each began, and no KEV entry yet tells agencies to rotate credentials.
- Inventory the keys each AI gateway, agent, MCP server and CI runner holds, and rehearse revoking all of them at once. Patching alone leaves leaked keys working.
CISA’s Known Exploited Vulnerabilities (KEV) catalog has added 11 entries in 2026 for tools that build, route and serve AI models and agents. In all of 2025 it added one. These tools sit where machine credentials collect, from model API keys to the cloud secrets an agent uses to reach other systems. When one is breached, the catalog asks agencies to apply the vendor’s fix. A patch closes the hole. It does not take back keys that already left.
What the catalog shows
We read the public mirror of the catalog (version 2026.11.13, 1,781 entries) with a short script and counted entries for five products: Langflow, LiteLLM, Ray, MLflow and GPUStack. Of the 11 entries from 2026, 6 gave agencies three days or fewer.
KEV entries for AI model and agent tools, by year added
| Product | 2025 | 2026 | Days to fix |
|---|---|---|---|
| Langflow | 1 | 5 | 3 to 14 |
| LiteLLM | 0 | 3 | 3 to 14 |
| GPUStack | 0 | 1 | 3 |
| MLflow | 0 | 1 | 14 |
| Ray | 0 | 1 | 3 |
| Total | 1 | 11 |
CISA KEV catalog version 2026.11.13, 1,781 entries. 2026 runs through November 13. Langflow includes IBM’s listing of it. We count by product name, so AI features inside larger products are not here. No entry for these five products predates 2025.
Read the descriptions and the theme is keys. LiteLLM proxies calls to model providers; its advisory says an attacker could reach the flaw through “any LLM API route.” CISA describes the result as “unauthorized access to the proxy and the credentials it manages.” A second LiteLLM entry let an attacker “establish an authenticated MCP session using an arbitrary Bearer token.” The GPUStack entry, added November 9, lists “persistent administrator API-key creation” and “disclosure of worker and cloud-provider secrets.” GPUStack’s advisory says the flaw was “observed exploited in the wild.”
In the pipeline, keys leave in hours
Build runners and developer tools hold credentials too, and three compromises with KEV entries show the timing.
- tj-actions/changed-files, March 2025. GitHub’s advisory says the malicious code “extracted secrets from the Runner Worker process memory and printed them in GitHub Actions logs,” affecting “over 23,000 repositories” between March 14 and 15. KEV listed it 4 days after it began, with a due date 25 days after.
- Trivy, March 2026. An attacker pushed “76 of 77 version tags” of an open-source vulnerability scanner’s GitHub Action to “credential-stealing malware.” Exposure windows ran from about 3 to about 12 hours. CISA says the code could reach “all tokens, SSH keys, cloud credentials, database passwords” in the pipeline. The KEV due date came 21 days after the March 19 compromise.
- TanStack, May 2026. “84 malicious versions across 42” packages were published “between approximately 19:20 and 19:26 UTC” on May 11, through “the legitimate GitHub Actions OIDC trusted-publisher binding.” The attacker took the short-lived OIDC token from runner memory. The KEV due date came 30 days after.
Across the three, the federal deadline landed 21 to 30 days after the compromise began. The credentials were exposed in the first hours. The Trivy advisory adds a harder lesson. After an earlier incident, rotation “was not atomic (not all credentials were revoked simultaneously),” and the attacker may have used a still-valid token to take the new secrets during a window “(which lasted a few days).” Read the advisory before you write your next rotation runbook.
Every agent is a credential holder
AI agents are adding holders quickly. GitHub now gives its Copilot cloud agent “its own dedicated Agents secrets and variables, alongside the existing Actions, Codespaces, and Dependabot types.” An organization can share one across “any or all repositories,” and the agent receives them “as environment variables.” The Model Context Protocol project’s guidance on local servers says a server “inherits your environment variables” and can read “SSH keys, cloud credentials, browser profiles.” It calls client configuration files “a known, predictable location on disk” and says to keep long-lived secrets out of them.
OWASP’s Non-Human Identities Top 10 rates secret leakage as easy to exploit and hard to detect, because “the secret enables the attacker to authenticate as the legitimate application.” To the logs, a stolen agent key looks like the agent doing its job.
Where the authorization package falls short
NIST SP 800-53 already covers this. IA-5(7) says to “ensure that unencrypted static authenticators are not embedded in applications or other forms of static storage.” IA-5 asks for procedures “for lost or compromised or damaged authenticators, and for revoking authenticators.” CM-8 asks for an inventory that “includes all components within the system.” AC-2 ties account notices to when “users are terminated or transferred.” An agent’s token has no HR record.
Our reading: an SSP that lists an AI gateway or an MCP server as a software component, without the keys it holds, cannot answer the first question an incident asks. Which credentials did this tool hold, what could each one reach, and who can revoke all of them at once?
Our forecast
- Our forecast: in calendar 2027, CISA will add at least 11 KEV entries for tools that build, route or serve AI models and agents, matching or passing 2026. We will rerun this count in January 2028 and publish it, with any new products named.
- Our forecast: by December 31, 2027, CISA will publish a KEV entry whose required action tells agencies to rotate or revoke credentials in addition to patching. As of version 2026.11.13, none of the 1,781 entries does. The catalog has changed shape once already: since June 25, 2026, 105 entries carry a forensic triage flag.
Both are falsifiable, and we will report back either way. The question for every ISSO, AO and program manager running an AI pilot is simpler: can you name every credential your AI agents and pipelines hold, and who would revoke all of them within an hour if one leaked?
What we’d do first
- Inventory credentials by component. For each AI gateway, agent, MCP server and CI runner, record the keys it holds, their scope, their owner, their expiry and the revoke path. Put it in the SSP under CM-8 and IA-5.
- Patch and rotate on the same day. When KEV lists a tool that holds credentials, treat it as a credential incident. Rotate everything the tool could reach, all at once, then verify the old keys fail.
- Give agents narrow, short-lived keys. One key per agent and per server, scoped to one job, with a human approving any step up. That is the human-in-the-loop model we describe in AI in government, governed.
- Make the blast radius visible. Our Next-Gen GRC beta is built around knowing the systems affected in seconds through the dependency graph. We would add credentials to that graph as edges, so a leaked key shows what it reaches, and let an unrotated key on an exposed AI tool move the AusperRiskIQ™ score. Evidence from pipelines keeps the record current between assessments; see our capabilities.
A note on method. Our script reads the pinned catalog and advisories, checks each quoted phrase, and reproduces every count here. For how short KEV deadlines have become under BOD 26-04, see our KEV clock analysis.
Common questions
What is a non-human identity?
Does patching a KEV-listed AI tool fix a credential leak?
Which NIST SP 800-53 controls apply?
Are short-lived tokens enough?
Sources
- CISA, Known Exploited Vulnerabilities catalog, version 2026.11.13 (JSON, public mirror cisagov/kev-data)Pinned commit of github.com/cisagov/kev-data. All KEV counts, dates and descriptions come from this file; make_ai_keys.py reproduces them. Accessed Nov 16, 2026.
- GitHub Advisory Database, GHSA-mrrh-fwg8-r2c3: tj-actions/changed-files (CVE-2025-30066)Accessed Nov 16, 2026.
- GitHub Advisory Database, GHSA-69fq-xp46-6x23: Trivy ecosystem supply chain compromise (March 2026)Accessed Nov 16, 2026.
- GitHub Advisory Database, GHSA-g7cv-rxg3-hmpx: malware in @tanstack/* packages (May 2026)Accessed Nov 16, 2026.
- GitHub Advisory Database, GHSA-r75f-5x8p-qvmc: LiteLLM SQL injection in proxy API key verificationAccessed Nov 16, 2026.
- GitHub Advisory Database, GHSA-gj9m-f8ph-pmhr: GPUStack authentication bypassAccessed Nov 16, 2026.
- GitHub Docs, Configure secrets and variables for Copilot cloud agentSource Markdown from github.com/github/docs, pinned commit. Accessed Nov 16, 2026.
- Model Context Protocol, Local Server Security (2026-07-28 docs)Source from github.com/modelcontextprotocol/modelcontextprotocol, pinned commit. Accessed Nov 16, 2026.
- OWASP Non-Human Identities Top 10 (2025), NHI2: Secret LeakageAccessed Nov 16, 2026.
- OWASP Non-Human Identities Top 10 (2025), IntroductionAccessed Nov 16, 2026.
- NIST SP 800-53 Rev. 5 catalog (OSCAL JSON, usnistgov/oscal-content)Control text for IA-5, IA-5(7), CM-8 and AC-2. Accessed Nov 16, 2026.
Share on LinkedInFollow AusperSubscribe by RSS
Related reading
Working on this now?
Tell us where your program stands and we'll tell you what we would do next.
Talk to Ausper