Attack ChainAIDerived
Hypothetical attack flow derived from CVE metadata – not a confirmed exploit chain
-
AccessAuthenticate to Omnigent server
-
ExploitUpload bundle with nested sub-agent escaping os_env.cwd
-
ExecutionBypass root-only workspace validation in bundles.py
-
ImpactRead/write host files as runner user
Vulnerability Assessment
| Weakness |
CWE
OWASP Top 10:2025
A01:2025 - Broken Access Control
ASVS 5.0
V5 - File Handling Path canonicalization, traversal prevention
MITRE ATT&CK
2 techniques across 2 tactics: Discovery, Collection The technique list with ATT&CK links and the kill-chain view is part of the Basic plan and up. |
| Exploitation Conditions | Requires an authenticated user who can upload an agent bundle to the Omnigent server (PR:L). The attacker-controlled bundle must define a nested sub-agent (or terminal) whose os_env.cwd is an absolute path or contains '..'; the pre-0.14.0 validation in bundles.py only checked the root agent's cwd, so nested specs bypassed it. The most severe host-filesystem read/write exposure arises on a runner that does NOT have OMNIGENT_RUNNER_WORKSPACE configured, where the supplied cwd becomes the environment root and copytree source (this is the AT:P attack requirement). Limiting factors: valid authentication plus bundle-upload rights are mandatory, exploitation is confined to the runner user's privileges, trusted local runs intentionally retain absolute-cwd behavior, and only versions 0.3.0 through 0.13.x are affected (fixed in 0.14.0). |
| Impact | This is a genuine but access-gated vulnerability rather than a mass-exploitation threat. The vendor CVSS 4.0 vector (7.7 High) shows AV:N and UI:N with full High impact on confidentiality, integrity and availability (VC:H/VI:H/VA:H), which looks severe, but two metrics temper it: PR:L means the attacker must already be an authenticated user with bundle-upload capability, and AT:P indicates a specific attack requirement - the code comments confirm the host-filesystem exposure is most acute on a runner without OMNIGENT_RUNNER_WORKSPACE set, where the uploaded cwd becomes the environment root and copytree source. No EPSS score, SSVC decision, KEV listing, or POC was provided in the input, so exploitation probability cannot be quantified; per the precision rules this is 'no public exploit identified at time of analysis.' The signals are internally consistent (network-reachable upload surface, authenticated low-priv actor, config-dependent host exposure), so there is no contradiction to flag - the main missing data are EPSS and any SSVC/automatability assessment. Real priority is moderate-to-high specifically for operators running Omnigent in multi-tenant or untrusted-bundle-upload mode; single-user or trusted-local deployments face little practical risk. |
| Exploit ScenarioDerived | An authenticated user with permission to upload an agent bundle crafts a bundle whose nested sub-agent sets os_env.cwd to an absolute path or one containing '..'. Because the pre-patch root-only check in bundles.py never recursed into sub-agents or terminals, the sub-agent's working directory escapes the workspace sandbox, letting the attacker read and write arbitrary host files as the runner user. No public exploit was identified in the provided data. |
Get an e-mail when this CVE is exploited, patched or re-scored. Free account.
Alert meAnalysisAIv1 – Oct 18, 2026
Path traversal in Omnigent 0.3.0 through 0.13.x allows authenticated users who can upload an agent bundle to escape the session workspace and read or write host files with the runner user's privileges by setting os_env.cwd to an absolute or parent-traversing path on a nested sub-agent or terminal. Validation in bundles.py only checked the root agent's working directory, so cwd values supplied by nested specs slipped through unvalidated. The most severe host-filesystem exposure arises on runners where OMNIGENT_RUNNER_WORKSPACE is not configured, because the attacker-supplied cwd then becomes the environment root and copytree source; no public exploit code identified at time of analysis, and exploitation requires valid credentials plus bundle-upload rights.
RemediationAI
Vendor-released patch: upgrade to Omnigent 0.14.0 (https://github.com/omnigent-ai/omnigent/releases/tag/v0.14.0), which contains commit 4a406d5b3b3f3a7c3401eab659ab04e728923b94 and replaces the root-only cwd check with recursive validation of the agent, its terminals, and all nested sub-agents. If an immediate upgrade is not possible, the single most effective compensating control is to set OMNIGENT_RUNNER_WORKSPACE on every runner to an explicit, non-sensitive workspace directory: this removes the condition where an attacker-supplied cwd becomes the environment root and copytree source, though it does not stop traversal relative to that workspace root and should be treated as a stopgap rather than a fix. Secondarily, restrict bundle-upload rights to trusted operators or require administrative review of uploaded bundles, inspecting any sub_agents and terminals entries for absolute or '..'-containing os_env.cwd values before acceptance; the trade-off is reduced self-service onboarding for agent authors. Running the Omnigent server and runner under a low-privilege account with no access to sensitive host paths limits the blast radius, since the traversal is confined to the runner user's permissions, but note that trusted local runs deliberately retain absolute-cwd behavior, so local execution is not a substitute for patching the upload path. Consult the vendor advisory at https://github.com/omnigent-ai/omnigent/security/advisories/GHSA-p5x3-3gqh-x44g for the authoritative fix guidance.
Affected VersionsENISA
Affected Packages
Source: OSV (Open Source Vulnerabilities)
Affects your stack
Affects 1 package in euvd
Checking this CVE against your own stacks and servers is part of the Basic plan and up.
Technical ContextAI · show details
The flaw is a classic CWE-22 path traversal rooted in how Omnigent materializes an uploaded agent definition. Omnigent agent bundles are YAML specs parsed into an AgentSpec model that includes an os_env object; before 0.14.0, the untrusted-upload validation path in omnigent/server/bundles.py (validate_agent_bundle, around lines 138-155) inspected only the top-level spec.os_env.cwd and rejected absolute paths or '..' segments there, but never recursed into spec.sub_agents or into per-entry spec.terminals os_env settings. Because sub-agents and terminals can each override the session working directory, an attacker could upload a bundle whose nested definition carried cwd: / or ../../.. and have that value accepted. On a runner without OMNIGENT_RUNNER_WORKSPACE set, the supplied cwd is used as the agent environment root and as the source for copytree, turning the traversal into arbitrary host file read/write under the runner account. The design intentionally preserves absolute-cwd behavior for trusted local runs (documented in designs/SESSION_WORKSPACE_SELECTION.md), so the trust boundary is specifically the untrusted bundle-upload path. The upstream fix commit 4a406d5b3b3f3a7c3401eab659ab04e728923b94 replaces the single root-only check with _reject_uploaded_escaping_cwd(), which walks the agent, every terminal's OSEnvSpec, and all sub-agents recursively via _reject_escaping_cwd(); the same commit also hardens policy-handler vetting in omnigent/inner/loader.py and omnigent/policies/registry.py by adding function_policy_handler_allowed(), which validates legacy handler/callable fields and shim-wrapped targets against the policy registry before parsing an uploaded policy. Impact is bounded by the runner user's file permissions and by the requirement for an authenticated account with bundle-upload capability.
Know when Omnigent gets its next vulnerability
- Watch products and get an email when something new is published
- Check this CVE against your own stack and servers
- Exploit and PoC references, threat intelligence from 22 sources
No credit card · 7-day full trial
Severity by source
CVSS:4.0/AV:N/AC:L/AT:P/PR:L/UI:N/VC:H/VI:H/VA:H/SC:N/SI:N/SA:N
Authenticated bundle upload gives PR:L; the runner-configuration dependency (no OMNIGENT_RUNNER_WORKSPACE) maps to AC:H in 3.1; host file read/write as runner yields C:H/I:H/A:H.
Primary rating from Vendor (VulnCheck).
CVSS VectorVendor: VulnCheck
DescriptionCVE.org
Omnigent 0.3.0 before 0.14.0 contains a path traversal vulnerability that allows authenticated users to escape the workspace by setting os_env.cwd on nested sub-agents. Attackers can upload a bundle whose sub-agent uses an absolute or parent-traversing working directory, bypassing root-only validation in bundles.py, to read and write host files as the runner user.
Technical POC Analysis
CWE-22 path traversal in Omnigent's untrusted bundle upload validation.
Full technical POC analysis - root cause, step-by-step exploitation and patch breakdown - is part of the Pro and Business plans.
Executive ViewAI
A path-traversal defect in the Omnigent agent platform (versions 0.3.0 through 0.13.x) allows accounts holding bundle-upload rights to step outside their assigned session workspace and read or alter files on the runner host using the runner service account's privileges, dissolving the isolation boundary between tenants and the infrastructure beneath them. Leadership should care because runners are typically shared: one crafted bundle can expose another customer's data or stage changes on production hosts, and the exposure is widest where OMNIGENT_RUNNER_WORKSPACE has not been configured, since the attacker-supplied working directory then becomes the environment root. No public exploit code identified at time of analysis, but a vendor patch is available and remediation effort is modest relative to the potential blast radius.
Apply vendor patch.
Threat Intelligence Overview
How this decision was reached
The engine reads each signal below and applies a fixed list of rules from top to bottom. The first rule that matches sets the recommended action – like a statement, every line is shown.
What every action level means
SSVC AssessmentENISA
Priority Score
Attack Technique
Path traversal exploits occur when applications use user-controlled input to construct file system paths without proper validation.
Remote code execution in Omnigent, an open-source AI agent framework and meta-harness, before version 0.3.0 allows an au
Remote code execution in Omnigent's multi-tenant runner allows any authenticated user to hijack the runner process by up
Path traversal in Omnigent AI agent framework (versions prior to 0.3.0) allows authenticated users to escape the session
Shell-command parser bypass in Omnigent's policy layer allows an authenticated or prompt-injected agent to push code to
EUVD-2026-103468
GHSA-vg9j-3mjv-5cq3
NVD
CVE.org
EPSS
Vulnrichment
Exploit-DB