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-2026-108595 Overview
CVE-2026-108595 is a permission bypass vulnerability in Phi versions 0.3.0 through 0.28.4. The flaw allows spawned sub-agents to escape workspace_only_writes and readonly mode by supplying an unchecked workdir to the agent_spawn tool. Attackers can embed prompt-injected instructions in processed content, causing the agent to spawn a worker rooted outside the sanctioned workspace. The resulting sub-agent inherits an attacker-controlled root, enabling unapproved file writes anywhere the invoking user can write. The weakness is tracked under CWE-863: Incorrect Authorization.
Critical Impact
Prompt-injected content can coerce Phi into spawning a sub-agent outside the workspace sandbox, resulting in arbitrary file writes across the user's filesystem.
Affected Products
- Phi 0.3.0 through 0.28.4 (PulseAI Club)
- Deployments configured with workspace_only_writes restrictions
- Deployments configured with readonly mode protections
Discovery Timeline
- 2026-10-10 - CVE-2026-108595 published to NVD
- 2026-10-16 - Last updated in NVD database
Technical Details for CVE-2026-108595
Vulnerability Analysis
Phi enforces sandbox restrictions through two modes: workspace_only_writes, which confines file writes to the configured workspace root, and readonly, which blocks writes entirely. These checks are evaluated against the parent agent's workspace path. The agent_spawn tool accepts a workdir parameter used to root a child worker. The vulnerable code paths in internal/agent/engine_runner.go and internal/tools/agenttool/agent.go do not validate that this workdir falls within the parent's permitted workspace. A sub-agent therefore runs with its own root and bypasses parent-level restrictions, performing writes that the parent could not. Technical details are documented in the VulnCheck Advisory on phi Permission Bypass and the HackMD Workspace Escape Analysis.
Root Cause
The root cause is missing authorization enforcement on the workdir argument passed to agent_spawn. Phi trusts the model-supplied path and applies no containment relative to the invoking agent's workspace. Because sandbox decisions are scoped per-agent rather than inherited across the spawn boundary, child agents can be instantiated with an arbitrary root. See the agent.go source and engine_runner.go source.
Attack Vector
An attacker plants prompt-injected instructions in content that Phi later processes, such as a document, repository file, web page, or ticket body. When the agent reads the poisoned content, the injected instructions direct it to call agent_spawn with a workdir outside the workspace. The spawned sub-agent then issues file write operations under the attacker-chosen directory. Exploitation requires user interaction with the Phi agent and attacker-controlled content in the processing pipeline, but no authentication to the host.
No verified exploit code is public. Refer to the external advisories
listed above for conceptual exploit flow and reproduction notes.
Detect Identity Attacks the Moment Authentication Fails
Singularity™ Identity surfaces credential misuse, lateral movement, and privilege escalation in real time, so you can contain the breach before it spreads.
Stop Identity AttacksDetection Methods for CVE-2026-108595
Indicators of Compromise
- Phi audit log entries showing agent_spawn invocations with a workdir outside the configured workspace root.
- Unexpected file creation or modification by the Phi process under user home directories, dotfile locations, or startup paths.
- Sub-agent sessions whose working directory differs from the parent agent's workspace path.
Detection Strategies
- Parse Phi engine logs for agent_spawn calls and alert when workdir is not a descendant of the parent workspace.
- Correlate agent process file-write telemetry with the configured workspace boundary to flag out-of-scope writes.
- Review content ingested by Phi for prompt-injection markers that reference agent_spawn, workdir, or sandbox-bypass language.
Monitoring Recommendations
- Enable file integrity monitoring on directories outside the Phi workspace that the service account can write to.
- Forward Phi runtime and tool-invocation logs to a centralized logging platform for retention and correlation.
- Monitor for new or unexpected child processes and file descriptors opened by Phi worker agents.
How to Mitigate CVE-2026-108595
Immediate Actions Required
- Upgrade Phi to a release later than 0.28.4 that enforces workspace containment on agent_spawnworkdir.
- Disable or restrict the agent_spawn tool in agent configurations until the upgrade is deployed.
- Run Phi under a low-privilege service account with filesystem ACLs that limit writable paths.
Patch Information
Monitor the PulseAI Club Phi repository and the VulnCheck advisory for the fixed release. The remediation must validate that any workdir passed to agent_spawn resolves within the parent agent's permitted workspace before a sub-agent is instantiated.
Workarounds
- Remove agent_spawn from the enabled tool set for untrusted or externally sourced content workflows.
- Treat all ingested documents, repositories, and web content as untrusted and sanitize before passing to Phi.
- Apply operating system mandatory access controls, such as AppArmor or SELinux, to constrain Phi's write surface.
- Run Phi inside a container with a read-only root filesystem and an explicit writable mount for the workspace only.
# Example container-level containment for Phi
docker run --rm \
--read-only \
--tmpfs /tmp:rw,noexec,nosuid,size=64m \
-v /srv/phi/workspace:/workspace:rw \
-u 1000:1000 \
phi:patched
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

