# The AI infrastructure playbook for the first 72 hours

Playbook, sources read 19 Nov 2026

**The short answer.** LiteLLM, Langflow and MLflow servers are often started by one engineer, left open, and loaded with model-provider keys and cloud roles. CISA listed flaws in all three between 4 Aug and 2 Sep 2026. Find the servers, block the exposed routes, upgrade, then rotate every key the tool stored.

Self-hosted AI gateways, workflow builders and experiment trackers behave like internal admin tools that were put on the internet for convenience. LiteLLM CVE-2026-59822 accepted a made-up Authorization header on its MCP endpoint and fell back to an empty auth object. Langflow CVE-2026-9198 let any network caller fetch a superuser token from an auto-login route and then run Python through a code-validation route. MLflow CVE-2026-64849 let an unauthenticated caller make the tracking server request internal addresses and read the answer. In all three the attacker needs no account, and the prize is what the server holds: provider API keys, stored secrets, tool access through MCP, and the cloud role of the host.

Patching is awkward for three reasons. The tools are rarely in the asset inventory, because a data scientist ran pip install on a notebook host. The fixes ship long before the CVE: the LiteLLM fix was released on 14 May 2026, 55 days before the NVD record, and the Langflow fix on 23 Jun 2026, 24 days before. And the fixes change behavior: LiteLLM 1.84.0 is flagged by the vendor as a release with breaking changes, and Langflow's switch away from auto-login can lock you out of an existing install's flows.

## The steps, hour by hour

| Window | Owner | Step, and why it comes here | Record to keep |
| --- | --- | --- | --- |
| 0 to 1 hour | Security lead | Find the servers, including the unregistered ones. On hosts: pip list 2>/dev/null \| grep -Ei 'litellm\|langflow\|mlflow'. In containers: docker ps --format '{{.Image}} {{.Ports}}' \| grep -Ei 'litellm\|langflow\|mlflow'. In Kubernetes: kubectl get deploy,sts -A -o wide \| grep -Ei 'litellm\|langflow\|mlflow'. From outside, scan your own public ranges for the ports these tools commonly use: Langflow's documentation uses 7860, LiteLLM's proxy commonly runs on 4000, MLflow's server on 5000. Ask the data and ML teams for the ones they started themselves.<br>These servers are not on the asset list, and the hour spent finding them is the hour the exposed ones stay open. | Per server: host or pod, tool, version, port, who started it, internet-reachable yes or no, and which cloud role it runs under. |
| 0 to 4 hours | Platform engineering | Check each server against the affected range and the precondition. LiteLLM below 1.84.0: pip show litellm, then grep -n -A6 mcp_servers config.yaml and grep -n -B2 -A8 allow_all_keys config.yaml; with no mcp_servers block there are no MCP tools to reach. Langflow 1.0.0 through 1.10.0: pip show langflow, then env \| grep LANGFLOW_AUTO_LOGIN and the .env file; treat unset on 1.10.0 or earlier as enabled. MLflow 3.3.0 up to but excluding 3.15.0: pip show mlflow, then ps -ef \| grep 'mlflow server' and read --backend-store-uri; webhooks need a SQL store, and the default SQLite store counts.<br>The version alone does not say whether you are exposed; the configuration decides, and it tells you which block to apply first. | Per server: version, the three precondition answers (config item present, auth state, store type) and exposed yes or no. |
| 0 to 4 hours | Platform engineering | Cut the reachable path now. LiteLLM: disable the MCP routes, or block /mcp/ and the related MCP paths (for example /mcp-rest/) at the reverse proxy, then check with curl -s -o /dev/null -w '%{http_code}\n' https://PROXY_HOST/mcp/ from outside, which must time out or be blocked. Langflow: put an authenticating reverse proxy in front, and set LANGFLOW_AUTO_LOGIN=false only after LANGFLOW_SUPERUSER_PASSWORD is set, or you lose access to the superuser's flows. MLflow: load the auth plugin and keep the server off untrusted networks. For all three, remove the public listener and allow only the networks that need it.<br>A route block takes minutes and stops the unauthenticated call, while the upgrade needs a test cycle. | The proxy rule or setting changed, the test result from outside, and the time it took effect. |
| 4 to 24 hours | Platform engineering | Upgrade: LiteLLM to 1.84.0 or later, after reading the release notes for breaking changes and testing clients in staging; Langflow to 1.10.1 or later, which stops issuing the 365-day superuser token and no longer executes code in validate_code; MLflow to 3.15.0 or later, and to 3.16.0 or later on IPv6 or NAT64 networks, because 3.15.0 left four IPv6-transition bypasses. Pin the version in the image tag or requirements file so a rebuild cannot go back.<br>The route block is a stopgap that breaks MCP clients; the upgrade is the actual fix, and IBM and the MLflow advisory list no workaround for Langflow or MLflow. | Version before and after, the pinned reference, and the staging test result for LiteLLM. |
| 4 to 24 hours | Security lead | Rotate everything the server stored. List the secrets: LiteLLM provider keys and issued virtual keys; Langflow global variables, API keys and every session issued under auto-login, including the 365-day superuser token named in the 1.10.1 release notes; MLflow artifact-store and database credentials. Turning auto-login off does not revoke earlier sessions or API keys, so revoke them by hand. Then ask each provider for its usage log from the exposure window and compare it with what you expect.<br>Whoever read the stored keys can use them from anywhere, and the patch does nothing about keys already copied. | Each secret rotated, with the time, the provider and the usage-log result. |
| 4 to 24 hours | Platform engineering | For MLflow, close the metadata path the SSRF reaches. From the server host run curl -s -m 3 -o /dev/null -w '%{http_code}\n' http://169.254.169.254/ to see whether it answers. On AWS require IMDSv2 and keep containers out: aws ec2 modify-instance-metadata-options --instance-id <id> --http-tokens required --http-put-response-hop-limit 1. Check the current setting with aws ec2 describe-instances --instance-ids <id> --query 'Reservations[].Instances[].MetadataOptions.HttpTokens'. Software that still uses IMDSv1 on the host will break, so list it first.<br>The MLflow bug returns response bodies from internal addresses, and metadata credentials are the most valuable thing there. | The metadata setting before and after, the software that broke, and the fix applied. |
| 4 to 24 hours | Security lead | Hunt in the proxy logs. Langflow: a request to /api/v1/auto_login followed by one to /api/v1/validate/code from the same source or from outside your network. LiteLLM: requests to /mcp/ or /mcp-rest/ from unknown addresses carrying an Authorization value that matches no key you issued; Wiz saw probes using bearer tokens of one character against model-listing endpoints. MLflow: POST to /api/2.0/mlflow/webhooks and /api/2.0/mlflow/webhooks/{id}/test, and a list of registered webhooks pointing at hosts you do not recognize. On the network: connections from the server to 169.254.169.254 or private ranges it has no reason to contact.<br>None of the three vendors published a log signature, so the route names and the odd bearer value are the only search terms there are. | The search text, the log range from the earliest affected release on the host, and each hit with source address and time. |
| 24 to 72 hours | Security lead | Look on the host. Run ps -eo pid,ppid,user,cmd --forest and look for the AI server process with a shell or an unfamiliar child. Check temporary paths for what Wiz found on its honeypots, which it did not tie to these CVEs: /tmp/x86_64 or /tmp/amd64, /tmp/.dbus-cache/gmon, a miner renamed unicorn under /app/data/.claude/, staging address 94.26.106[.]29 and 185.62.1[.]8, and the mining domain pool.hashvault[.]pro. For Langflow, CISA marks forensic triage as required, so copy the logs and take a memory image before you clean up.<br>A cryptominer is the usual payload on these servers, and it shows up as a process before it shows up as a bill. | The process list, any file found with its hash and timestamp, and the decision to clean or rebuild the host. |
| 24 to 72 hours | Cloud engineer | Shrink what the server's role can do. List the attached policies with aws iam list-attached-role-policies --role-name <role> and remove wildcard actions and any access to buckets or models the tool does not use. Deny outbound traffic from the host except to the model providers and stores it needs.<br>The next flaw in the tool will be exploited the same way, and a narrow role limits what it can reach. | The role before and after, and the egress rule set. |
| 24 to 72 hours | Application owner | Register each server with an owner and a release feed. Subscribe to the project's GitHub releases and treat any release note about auth, validation or code execution as a security fix. Put the server behind your single sign-on, then run the first check from the 0 to 4 hour window again to prove the route block can be lifted.<br>The fixes arrive weeks before the CVE, so the people who watch releases patch before the exploitation starts. | Owner name, the feed subscribed, the next review date. |

Print this table as the checklist: write the time each step was done in the record column.

## When no fix exists yet

Vendor workarounds in this class are rare. LiteLLM published one (disable MCP routes or block /mcp/ at the gateway). IBM lists no workaround for Langflow, and the MLflow advisory lists none. The controls that exist are the ones in the route-block step: an authenticating proxy, auth turned on, removing the listener from untrusted networks, IMDSv2 and egress limits. Each one costs: MCP clients and agents stop working behind a block, users sign in again after auth is on, and software on a host that uses IMDSv1 breaks.

Decide on offline by what the server holds and who can reach it. If the server is reachable from the internet, holds provider keys or a broad cloud role, and no authenticating proxy can go in front of it inside four hours, stop it. CISA's required action for the AI flaws it listed tells you to discontinue use of the product if mitigations are unavailable. Stopping a LiteLLM gateway or a Langflow editor costs the applications behind it, so name those before the decision, and record who made it and when.

## CVE-2026-9198 worked through, hour by hour

Langflow OSS 1.0.0 through 1.10.0 issued a superuser token to any caller of the auto-login route, and a code-validation route then executed Python at function-definition time. CVSS base score 9.8. The dated events, from the release notes, IBM, NVD and CISA, run for 42 days before the KEV listing. The sources give dates, not hours, so the table keeps days.

| When | What happened | Source |
| --- | --- | --- |
| 23 Jun 2026, Day 0 | Langflow 1.10.1 is released. The notes list 'do not execute code in validate_code' and 'stop issuing 365-day superuser token via auto_login', without naming a CVE. | [source](https://github.com/langflow-ai/langflow/releases/tag/v1.10.1) |
| 2 Jul 2026, Day 9 | IBM publishes its security bulletin, tells users to upgrade Langflow OSS to 1.10.1 and lists no workaround or mitigation. | [source](https://www.ibm.com/support/pages/node/7278927) |
| 17 Jul 2026, Day 24 | NVD publishes CVE-2026-9198. The CISA coordinator SSVC entry stamped the same day lists exploitation as active and the flaw as automatable. | [source](https://services.nvd.nist.gov/rest/json/cves/2.0?cveId=CVE-2026-9198) |
| 4 Aug 2026, Day 42 | CISA adds the CVE to KEV with forensic triage required and a due date of 7 Aug 2026. | [source](https://www.cisa.gov/known-exploited-vulnerabilities-catalog?field_cve=CVE-2026-9198) |
| 7 Aug 2026, Day 45 | KEV due date, three days after the listing. | [source](https://www.cisa.gov/known-exploited-vulnerabilities-catalog?field_cve=CVE-2026-9198) |
| 27 Aug 2026, Day 65 | Wiz publishes 90 days of honeypot telemetry, including attackers who compromised a Langflow honeypot and dropped cryptominers. Wiz does not tie those attacks to this CVE. | [source](https://www.wiz.io/blog/ai-infrastructure-honeypot) |

The playbook turns the 42 days into three. An operator who followed the release feed in the last step upgraded on 23 Jun and never saw the flaw. One who started the 72-hour playbook on the IBM bulletin of 2 Jul finished by 5 Jul, 12 days before the SSVC entry marked exploitation active and 30 days before KEV. One who waited for the KEV listing had 3 days to upgrade, rotate the 365-day superuser token and search for the auto_login and validate/code sequence, with forensic triage required. The same pattern held for the other two tools: LiteLLM's fix (v1.84.0, 14 May 2026) was out 111 days before its KEV listing on 2 Sep 2026, and MLflow 3.15.0 (31 Jul 2026) 19 days before its KEV listing on 19 Aug 2026.

## Related pages

- [The developer collaboration platform playbook for the first 72 hours](/playbooks/developer-collaboration-platform/)
- [The Linux server and SSH playbook for the first 72 hours](/playbooks/linux-server-ssh/)
- [Application runtime playbook: the first 72 hours when a flaw is in PHP or another language runtime on your servers](/playbooks/application-runtime/)

## Sources

- LiteLLM advisory GHSA-7488-6r32-c95q: https://github.com/BerriAI/litellm/security/advisories/GHSA-7488-6r32-c95q (read 19 Nov 2026)
- Wiz on AI infrastructure honeypot: https://www.wiz.io/blog/ai-infrastructure-honeypot (read 19 Nov 2026)
- IBM security bulletin for Langflow: https://www.ibm.com/support/pages/node/7278927 (read 19 Nov 2026)
- Langflow 1.10.1 release notes: https://github.com/langflow-ai/langflow/releases/tag/v1.10.1 (read 19 Nov 2026)
- MLflow advisory GHSA-7gwp-5pfp-969j: https://github.com/mlflow/mlflow/security/advisories/GHSA-7gwp-5pfp-969j (read 19 Nov 2026)
- LiteLLM v1.84.0 release: https://github.com/BerriAI/litellm/releases/tag/v1.84.0 (read 19 Nov 2026)
- MLflow v3.15.0 release: https://github.com/mlflow/mlflow/releases/tag/v3.15.0 (read 19 Nov 2026)
- CISA KEV entry for CVE-2026-9198: https://www.cisa.gov/known-exploited-vulnerabilities-catalog?field_cve=CVE-2026-9198 (read 19 Nov 2026)
- CISA KEV entry for CVE-2026-59822: https://www.cisa.gov/known-exploited-vulnerabilities-catalog?field_cve=CVE-2026-59822 (read 19 Nov 2026)
- CISA KEV entry for CVE-2026-64849: https://www.cisa.gov/known-exploited-vulnerabilities-catalog?field_cve=CVE-2026-64849 (read 19 Nov 2026)
- NVD record for CVE-2026-9198: https://services.nvd.nist.gov/rest/json/cves/2.0?cveId=CVE-2026-9198 (read 19 Nov 2026)
- NVD record for CVE-2026-59822: https://services.nvd.nist.gov/rest/json/cves/2.0?cveId=CVE-2026-59822 (read 19 Nov 2026)
- NVD record for CVE-2026-64849: https://services.nvd.nist.gov/rest/json/cves/2.0?cveId=CVE-2026-64849 (read 19 Nov 2026)
- Langflow documentation: API keys and authentication: https://docs.langflow.org/api-keys-and-authentication (read 19 Nov 2026)

Last reviewed 19 Nov 2026. Canonical: https://zerodayplaybook.com/playbooks/ai-infrastructure/
