AI Threat Alert
CVE-2026-10595: lollms: path traversal enables unauth file read
UNKNOWN CISA: TRACK*A path traversal flaw in parisneo/lollms 2.1.0 lets an unauthenticated attacker read arbitrary files on the server by sending URL-encoded dot-dot sequences that bypass Starlette's path normalization in the SPA catch-all route. lollms is a self-hosted LLM chat/orchestration UI, so a successful read can expose configuration files, API keys, session data, or model artifacts stored alongside the application, turning a single web request into a credential-harvesting or environment-mapping opportunity. There's no CVSS score published and EPSS sits at 0.49% (top 60th percentile), it isn't in CISA KEV, and no public exploit or Nuclei template exists yet, so near-term mass exploitation is unlikely; still, the bug requires zero authentication and zero user interaction, which keeps it attractive to opportunistic scanners once a PoC surfaces. Any team running lollms 2.1.0 or earlier, especially instances exposed beyond localhost, should upgrade to version 3 immediately and in the interim front the service with a reverse proxy that blocks encoded traversal sequences and restrict network exposure to trusted hosts only.
What is the risk?
Low-to-moderate near-term risk despite high technical impact: exploitation requires no authentication and a single crafted HTTP request, and the vulnerability class (CWE-23 path traversal) is trivial to exploit once understood, but current threat signals are muted. EPSS is 0.49% (top 60th percentile), CISA SSVC rates it TRACK_STAR (track over time, no immediate action), it is absent from CISA KEV, and no public exploit or Nuclei template has surfaced. The package carries a moderate risk score (36/100) with 11 other known CVEs, and lollms shows 0 tracked downstream dependents, which caps blast radius to organizations running the tool directly rather than as a transitive dependency. Risk should be reassessed upward quickly if a PoC or scanner template appears, given the trivial exploitation complexity.
How does the attack unfold?
What systems are affected?
| Package | Ecosystem | Vulnerable Range | Patched |
|---|---|---|---|
| LoLLMs | pip | — | No patch |
Do you use LoLLMs? You're affected.
How severe is it?
What should I do?
1 step-
Upgrade parisneo/lollms to version 3 or later, where the traversal handling is fixed. Until upgraded, do not expose the lollms UI directly to the internet; restrict access to localhost or an authenticated VPN/reverse-proxy, and add a reverse-proxy or WAF rule that rejects requests containing encoded traversal sequences (%2e%2e, ..%2f, and double-encoded variants) before they reach the app. For detection, monitor access logs for catch-all route requests containing %2e%2e or anomalous path depth, and use file-integrity monitoring to flag reads of sensitive paths (.env, credential stores, /etc/passwd). Rotate any API keys or credentials stored in files reachable by the lollms process, since no telemetry currently confirms whether pre-patch instances were exploited.
What does CISA's SSVC say?
Source: CISA Vulnrichment (SSVC v2.0). Decision based on the CISA Coordinator decision tree.
How is it classified?
Which compliance frameworks are affected?
This CVE is relevant to:
Frequently Asked Questions
What is CVE-2026-10595?
A path traversal flaw in parisneo/lollms 2.1.0 lets an unauthenticated attacker read arbitrary files on the server by sending URL-encoded dot-dot sequences that bypass Starlette's path normalization in the SPA catch-all route. lollms is a self-hosted LLM chat/orchestration UI, so a successful read can expose configuration files, API keys, session data, or model artifacts stored alongside the application, turning a single web request into a credential-harvesting or environment-mapping opportunity. There's no CVSS score published and EPSS sits at 0.49% (top 60th percentile), it isn't in CISA KEV, and no public exploit or Nuclei template exists yet, so near-term mass exploitation is unlikely; still, the bug requires zero authentication and zero user interaction, which keeps it attractive to opportunistic scanners once a PoC surfaces. Any team running lollms 2.1.0 or earlier, especially instances exposed beyond localhost, should upgrade to version 3 immediately and in the interim front the service with a reverse proxy that blocks encoded traversal sequences and restrict network exposure to trusted hosts only.
Is CVE-2026-10595 actively exploited?
No confirmed active exploitation of CVE-2026-10595 has been reported, but organizations should still patch proactively.
How to fix CVE-2026-10595?
Upgrade parisneo/lollms to version 3 or later, where the traversal handling is fixed. Until upgraded, do not expose the lollms UI directly to the internet; restrict access to localhost or an authenticated VPN/reverse-proxy, and add a reverse-proxy or WAF rule that rejects requests containing encoded traversal sequences (%2e%2e, ..%2f, and double-encoded variants) before they reach the app. For detection, monitor access logs for catch-all route requests containing %2e%2e or anomalous path depth, and use file-integrity monitoring to flag reads of sensitive paths (.env, credential stores, /etc/passwd). Rotate any API keys or credentials stored in files reachable by the lollms process, since no telemetry currently confirms whether pre-patch instances were exploited.
What systems are affected by CVE-2026-10595?
This vulnerability affects the following AI/ML architecture patterns: self-hosted LLM chat UIs, model serving, local inference deployments.
What is the CVSS score for CVE-2026-10595?
No CVSS score has been assigned yet.
What is the AI security impact?
Affected AI Architectures
MITRE ATLAS Techniques
AML.T0007 Discover AI Artifacts AML.T0037 Data from Local System AML.T0049 Exploit Public-Facing Application Compliance Controls Affected
What are the technical details?
Original Advisory
A path traversal vulnerability exists in parisneo/lollms version 2.1.0, specifically in the SPA catch-all route implemented in `backend/routers/ui.py`. The vulnerability arises from the improper handling of user-controlled path input, which is directly joined into a filesystem path without sanitization or containment checks. URL-encoded dot-dot sequences (`%2e%2e`) bypass Starlette's built-in path normalization and are resolved by Python's `pathlib`, allowing an unauthenticated attacker to read arbitrary files on the server. This issue has been resolved in version 3.
Exploitation Scenario
An attacker discovers an internet- or network-exposed lollms instance (e.g., via a port scan or during recon of a target's internal AI tooling) and sends an HTTP GET to the SPA catch-all route with a URL-encoded traversal payload targeting files outside the intended web root. Because Starlette's normalization does not catch %2e%2e before pathlib resolves it, the request returns file contents directly in the response with no authentication required. The attacker uses this to pull .env or config files containing API keys for the LLM providers connected to lollms, then pivots to abuse those credentials for further access or cost harvesting, or reads local persona/RAG files to map and later target the deployment.
Weaknesses (CWE)
CWE-23 — Relative Path Traversal: The product uses external input to construct a pathname that should be within a restricted directory, but it does not properly neutralize sequences such as ".." that can resolve to a location that is outside of that directory.
- [Implementation] Assume all input is malicious. Use an "accept known good" input validation strategy, i.e., use a list of acceptable inputs that strictly conform to specifications. Reject any input that does not strictly conform to specifications, or transform it into something that does. When performing input validation, consider all potentially relevant properties, including length, type of input, the full range of acceptable values, missing or extra inputs, syntax, consistency across related fields, and conformance to business rules. As an example of business rule logic, "boat" may be syntactically valid because it only contains alphanumeric characters, but it is not valid if the input is only expected to contain colors such as "red" or "blue." Do not rely exclusively on looking for malicious or malformed inputs. This is likely to miss at least one undesirable input, especially if the code's environment changes. This can give attackers enough room to bypass the intended validation. However, denylis
- [Implementation] Inputs should be decoded and canonicalized to the application's current internal representation before being validated (CWE-180). Make sure that the application does not decode the same input twice (CWE-174). Such errors could be used to bypass allowlist validation schemes by introducing dangerous inputs after they have been checked. Use a built-in path canonicalization function (such as realpath() in C) that produces the canonical version of the pathname, which effectively removes ".." sequences and symbolic links (CWE-23, CWE-59). This includes: realpath() in C getCanonicalPath() in Java GetFullPath() in ASP.NET realpath() or abs_path() in Perl realpath() in PHP
Source: MITRE CWE corpus.
References
Timeline
Related Vulnerabilities
CVE-2026-1114 9.8 lollms: weak JWT secret allows admin takeover
Same package: lollms CVE-2026-1115 9.6 lollms: Stored XSS enables wormable account takeover
Same package: lollms CVE-2024-6982 8.4 lollms: RCE via eval() sandbox bypass in Calculate
Same package: lollms CVE-2026-1117 8.2 lollms: Access Control bypass enables privilege escalation
Same package: lollms CVE-2025-6386 7.5 lollms: timing attack enables credential enumeration
Same package: lollms