Disclaimer: This content was generated using AI. While we strive for accuracy, please verify critical information with official sources.
Sources: National Vulnerability Database, CISA KEV, Exploit-DB
CVE-2026-77294 Overview
TREK is a collaborative travel planner that integrates large language model (LLM) functionality for AI-assisted trip import. Versions prior to 3.3.0 allow an authenticated user with write access to a trip instance to store an attacker-controlled llm_base_url through the settings API when the LLM_PARSING feature is enabled. The stored URL is consumed by LLM client code without applying the server-side request forgery (SSRF) guard, enabling the server to issue outbound requests to attacker-chosen destinations. This vulnerability is tracked as [CWE-918] and is fixed in TREK version 3.3.0.
Critical Impact
An authenticated attacker can coerce the TREK server into requesting internal services and link-local cloud metadata endpoints, potentially disclosing infrastructure credentials and enabling modification of protected cloud resources.
Affected Products
- TREK collaborative travel planner, all versions prior to 3.3.0
- Deployments where the LLM_PARSING feature flag is enabled
- Instances exposing the settings API to authenticated users with trip write permissions
Discovery Timeline
- 2026-09-24 - CVE-2026-77294 published to NVD
- 2026-09-25 - Last updated in NVD database
Technical Details for CVE-2026-77294
Vulnerability Analysis
The vulnerability is a server-side request forgery (SSRF) in the AI-assisted trip import feature. TREK permits authenticated users to configure an LLM endpoint via the llm_base_url setting. When AI-assisted parsing is triggered, the server dispatches HTTP requests to that URL using one of three LLM clients: server/src/nest/llm-parse/clients/openai-compatible.client.ts, server/src/nest/llm-parse/clients/anthropic.client.ts, or server/src/nest/llm-parse/router/ollama-format.client.ts.
None of these client paths apply the server-side SSRF guard before issuing the outbound request. The upstream error response body is reflected back to the caller through parsing warnings, which creates a usable response channel for the attacker. This permits reconnaissance of internal services and retrieval of cloud instance metadata.
Root Cause
The root cause is the omission of the SSRF validation routine in the three LLM client modules. The destination URL is treated as trusted configuration rather than attacker-influenced input, so loopback, private, and link-local addresses are not blocked prior to the outbound fetch.
Attack Vector
Exploitation requires an authenticated account with write permission to a target trip instance and an environment where LLM_PARSING is enabled. The attacker stores a malicious llm_base_url through the settings API, pointing to an internal address such as a cloud metadata service endpoint. Triggering AI-assisted trip parsing then causes the TREK server to issue a request to that destination. Error text returned from the forged request is surfaced in parsing warnings, enabling credential and metadata disclosure that can be chained to modify protected cloud resources.
No verified public exploit code is available for this issue. See the GitHub Security Advisory GHSA-fmq9-ggh3-647p for maintainer details.
Detection Methods for CVE-2026-77294
Indicators of Compromise
- Settings API writes that set llm_base_url to loopback (127.0.0.1), private RFC1918 ranges, or link-local addresses such as 169.254.169.254.
- Outbound HTTP requests from the TREK server process to cloud metadata endpoints or internal-only hostnames.
- Parsing warning messages in application logs that contain upstream response bodies from unexpected hosts.
Detection Strategies
- Inspect application logs for llm_base_url configuration changes and correlate them with subsequent AI-assisted trip parsing events.
- Monitor the TREK server's egress traffic for connections to 169.254.169.254, metadata.google.internal, and other provider-specific metadata endpoints.
- Review authentication logs for accounts that modified LLM-related settings shortly before triggering trip imports.
Monitoring Recommendations
- Enable network flow logging on hosts running TREK and alert on outbound requests to metadata or loopback destinations.
- Capture and retain full request/response bodies for the settings API to support post-incident investigation.
- Alert on repeated LLM_PARSING failures that return non-LLM response payloads, which indicate probing of internal services.
How to Mitigate CVE-2026-77294
Immediate Actions Required
- Upgrade TREK to version 3.3.0 or later, which applies the SSRF guard to the affected LLM client paths.
- Audit the llm_base_url values currently stored for every trip and reset any entries pointing to internal, loopback, or link-local addresses.
- Revoke and rotate any cloud credentials reachable from the TREK server's instance metadata endpoint if exposure is suspected.
Patch Information
The fix is included in TREK release v3.3.0, delivered through Pull Request #1520 and commit 19064b3. The patch applies SSRF validation to the OpenAI-compatible, Anthropic, and Ollama-format LLM clients. Representative changes from the release commit:
<link rel="preconnect" href="https://fonts.googleapis.com" />
<link rel="preconnect" href="https://fonts.gstatic.com" crossorigin />
<link href="https://fonts.googleapis.com/css2?family=MuseoModerno:wght@400;700;800&display=swap" rel="stylesheet" />
-
- <!-- Leaflet -->
- <link rel="stylesheet" href="https://unpkg.com/leaflet@1.9.4/dist/leaflet.css"
- integrity="sha256-p4NxAoJBhIIN+hmNHrzRCf9tD/miZyoHS5obTRR9BMY="
- crossorigin="" />
</head>
<body>
<div id="root"></div>
Source: GitHub Commit 19064b3
Workarounds
- Disable the LLM_PARSING feature flag until the upgrade is applied to remove the vulnerable code path entirely.
- Restrict the TREK server's egress at the network layer, blocking access to 169.254.169.254 and other metadata or internal-only destinations.
- Limit which authenticated accounts hold write permission on trip instances to reduce the pool of users able to reach the vulnerable path.
# Example egress block for AWS metadata service (iptables)
iptables -A OUTPUT -d 169.254.169.254 -p tcp --dport 80 -j REJECT
iptables -A OUTPUT -d 169.254.169.254 -p tcp --dport 443 -j REJECT
Disclaimer: This content was generated using AI. While we strive for accuracy, please verify critical information with official sources.
Sources: National Vulnerability Database, CISA KEV, Exploit-DB

