# CVE-2026-59822: BerriAI LiteLLM, is it exploitable where you run it

Ledger entry, disclosed 8 Jul 2026, sources read 19 Nov 2026

**The short answer.** CVE-2026-59822 affects BerriAI LiteLLM and was disclosed 8 Jul 2026 with a CVSS score of 8.8. Exposure needs these conditions: the proxy runs LiteLLM earlier than 1.84.0; an MCP server in config.yaml sets allow_all_keys: true, and /mcp/ is reachable by the attacker. CISA added it to the Known Exploited Vulnerabilities catalog on 2 Sep 2026. Exploited 50 days after disclosure.

## What the flaw is

LiteLLM is a proxy that routes calls to LLM providers. It can also front MCP (Model Context Protocol) tool servers. CVE-2026-59822 is an improper authentication flaw (CWE-287) in the proxy's MCP Streamable HTTP endpoint. The MCP auth handler supports an OAuth2 passthrough mode for upstream MCP servers. In affected versions, when LiteLLM key validation failed, the fallback path accepted the request anyway and replaced the failed result with an empty auth object.

An attacker who can reach the proxy sends a made-up Authorization header and gets an authenticated MCP session without a valid LiteLLM key. From there the attacker can list and call the MCP tools the proxy has configured and reach the services behind them. The fix pull request says the exposure covers MCP servers marked allow_all_keys, because those servers are available to every caller. Wiz notes that a proxy often holds model-provider keys and cloud permissions, so the impact can extend well past the proxy.

Ratings: the CNA (GitHub) rates it 8.8 under CVSS 4.0 (network, low complexity, no privileges, high confidentiality impact, low integrity impact). NVD's own CVSS 3.1 base score is 8.2 (CVSS:3.1/AV:N/AC:L/PR:N/UI:N/S:U/C:H/I:L/A:N); the two scales differ, so the numbers are not in conflict.

## Are you exposed: each condition and how to check it

| Condition for exposure | How to check it on your system |
| --- | --- |
| the proxy runs LiteLLM earlier than 1.84.0 | On the host run pip show litellm, or pip list \| grep -i litellm. For a container run docker exec CONTAINER pip show litellm. 1.84.0 is the first fixed release (GitHub release v1.84.0). GitHub's advisory range is every version below 1.84.0 and names no start; the versions table gives the first build that carries the OAuth2 fallback the advisory describes. |
| an MCP server in config.yaml sets allow_all_keys: true, and /mcp/ is reachable by the attacker | Exposed servers: the fix pull request #26463 says the failed-key fallback returned an empty auth object whose tool access came from every server marked allow_all_keys: true, so any server with that setting is exposed on 1.83 and earlier. Run grep -n -A8 mcp_servers config.yaml (the --config path), then grep -n -B2 -A8 allow_all_keys config.yaml. A proxy with no MCP servers has no MCP tools to reach, and the pull request describes no access to a server without allow_all_keys. Reachability: from a network position an attacker could use, run curl -s -o /dev/null -w '%{http_code}\n' https://PROXY_HOST/mcp/. A 401 or 403 means the route is reachable and rejects anonymous callers; a timeout or block means the gateway already filters it. Also review gateway rules for /mcp/ and related MCP paths. Network control: check the bind address and firewall or security group rules for the proxy port with ss -ltnp \| grep litellm, then your firewall or cloud security group listing. The proxy should not accept connections from untrusted networks. |

## Affected and fixed versions

| Product line | Affected | First fixed | Notes |
| --- | --- | --- | --- |
| LiteLLM (pip package litellm, proxy server) | 1.81.9 to 1.83.x | 1.84.0 (released 14 May 2026) | 1.81.9 is the first release with the OAuth2 fallback described in the advisory (nightly build of 7 Feb 2026, stable of 15 Feb 2026). The fix pull request #26463 was merged on 30 Apr 2026. |
| LiteLLM (pip package litellm, proxy server) | Before 1.81.9 | 1.84.0 | Inside GitHub's range (below 1.84.0), but the OAuth2 fallback code is absent. The fix also closes a second route to the same empty auth object: the handler tests the full request URL, query string included, for the text .well-known. That test is present from 1.77.7 on, and the allow_all_keys server setting exists by 1.80.13. Upgrade rather than rely on the older code. |

The GitHub release page for v1.84.0 states that the release contains breaking changes and points to the v1.84.0 release notes on docs.litellm.ai. Read them before an upgrade from an older line.

## If the fix cannot go in yet

- Upgrade to LiteLLM 1.84.0 or later. This is the fix. Cost: the vendor flags 1.84.0 as a release with breaking changes, so test configuration and clients first.
- If you cannot upgrade yet, disable MCP routes, or block /mcp/ and related MCP endpoints at the reverse proxy or API gateway (GitHub advisory workaround). Cost: every MCP client and agent that reaches tools through this proxy stops working until the route is restored.
- Keep the proxy off untrusted networks, scope its cloud IAM permissions narrowly and restrict its outbound network access (Wiz guidance). Cost: tools or providers that rely on broad egress or permissions need explicit allow rules.
- CISA's KEV entry gives this required action: apply mitigations in accordance with vendor instructions, follow the BOD 26-04 guidance the entry cites, and discontinue use of the product if mitigations are unavailable. The KEV due date was 16 Sep 2026.

## What to look for

- Neither the vendor nor Wiz published a log signature for this CVE. The practical review: in proxy and gateway access logs, look for requests to /mcp/ and the MCP REST paths (for example /mcp-rest/) from addresses you do not know, especially ones carrying an Authorization header that does not match any key you issued, and tool listing or tool call activity that no known client explains.
- Wiz says its honeypots saw single-character bearer tokens (Authorization: Bearer x), with GET /v1/models as its printed example. A burst of requests with tiny bearer values from one source, to /mcp/ or to a model route, matches that pattern.
- Wiz guidance: watch the runtime for an AI server process spawning a shell or other unexpected child processes.
- A separate LiteLLM flaw, CVE-2026-42271, has its own Wiz indicators (download server 185.62.1[.]8, staging directory /tmp/.dbus-cache/). Finding them means that other flaw was used and does not confirm this one.

## Timeline

| Date | Event | Source |
| --- | --- | --- |
| 6 Feb 2026 | Commit 6a213fc (pull request #20602, OAuth2 passthrough for upstream MCP servers) adds the fallback that returns an empty auth object when LiteLLM key validation fails. It ships in the 1.81.9 nightly build of 7 Feb 2026. | [source](https://github.com/BerriAI/litellm/commit/6a213fc3bc) |
| 30 Apr 2026 | The fix, pull request #26463 (tighten public-route detection and OAuth2 fallback gating), is merged. | [source](https://github.com/BerriAI/litellm/pull/26463) |
| 14 May 2026 | LiteLLM v1.84.0 is released; it carries the fix. | [source](https://github.com/BerriAI/litellm/releases/tag/v1.84.0) |
| 8 Jul 2026 | NVD publishes CVE-2026-59822 (source: GitHub). | [source](https://services.nvd.nist.gov/rest/json/cves/2.0?cveId=CVE-2026-59822) |
| 22 Jul 2026 | GitHub publishes the reviewed advisory GHSA-7488-6r32-c95q with CVSS 4.0 score 8.8. | [source](https://api.github.com/advisories?cve_id=CVE-2026-59822) |
| 27 Aug 2026 | Wiz publishes its 90-day honeypot report and says it saw this bug exploited; the report gives no first day, so exploitation began on or before this date. | [source](https://www.wiz.io/blog/ai-infrastructure-honeypot) |
| 2 Sep 2026 | CISA adds the CVE to KEV with a due date of 16 Sep 2026. | [source](https://www.cisa.gov/known-exploited-vulnerabilities-catalog?field_cve=CVE-2026-59822) |

Exploited 50 days after disclosure: disclosed 8 Jul 2026, first exploitation reported 27 Aug 2026.

## Exploitation reported so far

Wiz Research published a report on 27 Aug 2026 covering 90 days of honeypot telemetry. It says it observed exploitation of CVE-2026-59822 against its LiteLLM honeypots, with requests that carried a one-character bearer token (Authorization: Bearer x). The one request Wiz prints is GET /v1/models, a model-listing route, not a request to /mcp/. Wiz describes the flaw as granting full MCP access to any bearer token and prints no /mcp/ request, so the report ties its honeypot traffic to this CVE by Wiz's own attribution. The report gives no first day for this bug, so the 50 days from the 8 Jul 2026 disclosure to the Wiz report shown on this page is a ceiling: use began on or before 27 Aug 2026 and may have begun earlier. The fix shipped in 1.84.0 on 14 May 2026, 55 days before the CVE record of 8 Jul 2026, and GitHub published the reviewed advisory on 22 Jul 2026.

CISA added the CVE to KEV on 2 Sep 2026, six days after the Wiz report. KEV lists ransomware campaign use as Unknown. The same Wiz report ties a different LiteLLM flaw (CVE-2026-42271, a command injection in the MCP server test endpoints) to a Qilin-linked chain. That claim does not apply to CVE-2026-59822.

## Why runtime reachability settles it

Exposure turns on three facts that only the running proxy shows together: a version below 1.84.0, an MCP server with allow_all_keys: true, and an MCP route an outsider can reach. KEV and Wiz confirm real attacks, so treat a reachable proxy as urgent. The checks table settles it row by row. See [runtime reachability](/methods/runtime-reachability/) for what the method proves and misses.

## Related pages

- [CVE-2026-9198: IBM Langflow OSS](/ledger/cve-2026-9198/)
- [CVE-2026-16723: Alibaba fastjson 1.x](/ledger/cve-2026-16723/)
- [CVE-2026-53266: Linux kernel](/ledger/cve-2026-53266/)

## Sources

- CVE-2026-59822: NVD record: https://services.nvd.nist.gov/rest/json/cves/2.0?cveId=CVE-2026-59822 (read 19 Nov 2026)
- CVE-2026-59822: CISA KEV entry (feed version 2026.11.18): https://www.cisa.gov/known-exploited-vulnerabilities-catalog?field_cve=CVE-2026-59822 (read 19 Nov 2026)
- CVE-2026-59822: vendor advisory (GitHub security advisory GHSA-7488-6r32-c95q, read through the GitHub advisory API record): https://github.com/BerriAI/litellm/security/advisories/GHSA-7488-6r32-c95q (read 19 Nov 2026)
- CVE-2026-59822: GitHub advisory API record (publish date, CVSS 4.0, credits): https://api.github.com/advisories?cve_id=CVE-2026-59822 (read 19 Nov 2026)
- CVE-2026-59822: Wiz Research honeypot report: https://www.wiz.io/blog/ai-infrastructure-honeypot (read 19 Nov 2026)
- CVE-2026-59822: vendor release page: https://github.com/BerriAI/litellm/releases/tag/v1.84.0 (read 19 Nov 2026)
- CVE-2026-59822: vendor fix pull request: https://github.com/BerriAI/litellm/pull/26463 (read 19 Nov 2026)
- CVE-2026-59822: vendor MCP documentation: https://docs.litellm.ai/docs/mcp (read 19 Nov 2026)
- CVE-2026-59822: vendor commit that adds the OAuth2 fallback (first shipped in 1.81.9): https://github.com/BerriAI/litellm/commit/6a213fc3bc (read 19 Nov 2026)

Last reviewed 19 Nov 2026. Canonical: https://reachabilityanalysis.com/ledger/cve-2026-59822/
