Skip to navigation

Threat reportVulnerabilityTL-2026-1921

Odysseus AI Workspace Remote Code Execution via Authorization Bypass — GHSA-xwhc-f36c-v5vm (CVSS 9.9)

criticalPATCHED

A critical vulnerability in Odysseus, a privacy-focused AI workspace for LLM interaction, allows any authenticated non-admin user to execute arbitrary OS commands with the privileges of the Odysseus process by smuggling an admin-only shell action onto a scheduled task across two ordinary API requests.

Key facts

IdentifiersTL-2026-1921, CVE-2026-70619, CVE-2026-70620
SeverityCRITICAL
CVSS9.9 (CVSS:3.1/AV:N/AC:L/PR:L/UI:N/S:C/C:H/I:H/A:H)
Exploitation statusPublic proof-of-concept exploit for a linked CVE
EPSSCVE-2026-70619: 0.3% (29.0th percentile); CVE-2026-70620: 0.2% (18.0th percentile)
Affected productsOdysseus AI Workspace
Weaknesses (CWE)CWE-863, CWE-306
MalwarePewDiePie, Python
MITRE ATT&CK techniquesT1005 Data from Local System, T1021 Remote Services, T1048 Exfiltration Over Alternative Protocol, T1053 Scheduled Task/Job, T1057 Process Discovery, T1059 Command and Scripting Interpreter, T1071 Application Layer Protocol, T1078 Valid Accounts, T1082 System Information Discovery, T1548 Abuse Elevation Control Mechanism, T1552 Unsecured Credentials, T1555 Credentials from Password Stores
MotivationUNKNOWN
Target sectorstechnology, open-source-software, individuals, startups
Target regionsGlobal
Detection rules9 (1 shown free)
Sources12 cited sources

Detection coverage

As of 2026-08-06, Threadlinqs Intelligence publishes 9 detection rule(s) for TL-2026-1921 across Splunk SPL, Microsoft KQL and Sigma, covering 6 indicator(s) of compromise. The whole corpus is readable without an account. Every threat comes with one complete detection in Splunk SPL, Microsoft KQL and Sigma, free for everyone; a free account adds daily TL Creds to unlock the rest. Paid plans add a larger daily allowance, correlation and the MCP server. Threadlinqs MCP server · View plans.

Free detection: Odysseus Process Spawning Shell Subprocess — Post-Exploitation Command Execution

Detects the Odysseus application process (Python/FastAPI) spawning a Unix shell (/bin/sh, /bin/bash) or executing commands via subprocess.run(..., shell=True). After successful authorization bypass, the attacker's injected payload runs as the odysseus user (uid 1000 in Docker) with no sandboxing. This detection monitors auditd or syslog for parent-child process relationships where parent is Python (odysseus process) and child is sh/bash executing commands. Also covers child processes like curl, wget, openssl for data exfiltration, and access to sensitive files (password hashes, TOTP secrets, SSH keys).

  • Languages: Splunk SPL, Microsoft KQL, Sigma
  • ATT&CK: T1059, T1005, T1552, T1057, T1082
  • Data sources: linux_auditd, sysmon_linux, process_creation_events, linux_syslog
  • Severity: critical
  • Confidence: high

Splunk SPL

index=audit sourcetype=linux_audit
  (auid>=1000 OR auid=4294967295)
  (type=EXECVE OR type=SYSCALL)
| search (comm IN ("/bin/sh", "/bin/bash", "sh", "bash") AND ppid>0)
  OR (exe IN ("/bin/sh", "/bin/bash", "/usr/bin/sh", "/usr/bin/bash"))
| eval parent_cmd=if(isnull(ppid), "unknown", ppid)
| lookup process_lookup pid as ppid OUTPUT cmdline as parent_cmdline
| where parent_cmdline LIKE "%odysseus%" OR parent_cmdline LIKE "%python%" OR parent_cmdline LIKE "%uvicorn%"
| eval cmdline_lower=lower(ARGV)
| eval suspicious_cmds=if(
    match(cmdline_lower, "(curl|wget|nc|ncat|bash -i|sh -i|openssl|base64 -d|chmod +x|/dev/tcp)")
    OR match(cmdline_lower, "(passwd|shadow|id_rsa|id_ed25519|authorized_keys|.ssh)")
    OR match(cmdline_lower, "(export.*TOKEN|export.*API_KEY|export.*SECRET|env |printenv)")
  , 1, 0)
| where suspicious_cmds=1
| table _time, host, uid, ppid, pid, exe, ARGV, parent_cmdline, uid, session
| eval alert_severity=if(
    match(ARGV, "(id_rsa|passwd|shadow|export.*API_KEY|TOTP|secrets)"), "critical",
    "high"
  )
| sort -_time
| rename ARGV as command_line, exe as executable_path, uid as user_id

Microsoft KQL

DeviceProcessEvents
| where Timestamp > ago(7d)
| where InitiatingProcessFileName in~ ("python3", "python", "uvicorn", "odysseus")
| where FileName in~ ("sh", "bash", "dash", "zsh")
| extend CmdLine = iff(isempty(ProcessCommandLine), CommandLine, ProcessCommandLine)
| where CmdLine contains "-c" or CmdLine contains "shell=True" or CmdLine contains "subprocess"
| extend SuspiciousIndicator = case(
    CmdLine has_any ("curl", "wget", "nc", "ncat", "bash -i", "sh -i"), "data_exfil_or_rev_shell",
    CmdLine has_any ("passwd", "shadow", "id_rsa", "authorized_keys", ".ssh"), "credential_access",
    CmdLine has_any ("export", "env", "printenv"), "credential_enumeration",
    CmdLine has_any ("chmod +x", "chmod 777"), "binary_staging",
    CmdLine has_any ("/dev/tcp", "/dev/udp", "openssl"), "network_backdoor",
    CmdLine has_any ("base64", "xxd", "strings"), "encoding_evasion",
    "other"
  )
| where SuspiciousIndicator != "other"
| project Timestamp, DeviceName, InitiatingProcessFileName, FileName, CmdLine, SuspiciousIndicator, AccountName
| extend AlertSeverity = iff(SuspiciousIndicator in~ ("credential_access", "network_backdoor", "rev_shell"), "critical", "high")
| order by Timestamp desc

Sigma

title: Odysseus Process Spawning Shell Subprocess — Post-Exploitation Command Execution
id: d7f1b3c2-9e8a-4f16-8d52-7c3a9b0e1d4f
status: experimental
description: Detects the Odysseus application process (Python) spawning a Unix shell interpreter as a child process, indicating post-exploitation command execution via the scheduled task authorization bypass. Monitors for suspicious child process arguments including data exfiltration tools, credential access commands, and reverse shell patterns.
references:
  - https://www.manifold.security/blog/odysseus-ai-workspace-rce
  - https://attack.mitre.org/techniques/T1059/004/
  - https://attack.mitre.org/techniques/T1005/
author: Threadlinqs Team
date: 2026-08-06
tags:
  - attack.execution
  - attack.collection
  - attack.credential_access
  - attack.t1059
  - attack.t1005
  - attack.t1552
  - attack.t1057
  - attack.t1082
  - odysseus.rce
  - subprocess.shell_execution
logsource:
  category: process_creation
  product: linux
  service: auditd
  product_optional: sysmon
  service_optional: sysmon_linux
detection:
  selection_parent:
    ParentImage|endswith:
      - '/python3'
      - '/python'
      - '/uvicorn'
      - '/odysseus'
  selection_child:
    Image|endswith:
      - '/sh'
      - '/bash'
      - '/dash'
      - '/zsh'
  selection_suspicious_args:
    CommandLine|contains:
      - '-c'
      - 'curl'
      - 'wget'
      - 'passwd'
      - 'shadow'
      - 'id_rsa'
      - 'authorized_keys'
      - '.ssh'
      - 'chmod +x'
      - '/dev/tcp'
      - 'base64 -d'
      - 'openssl'
      - 'export '
      - 'env |'
  condition: selection_parent and selection_child and selection_suspicious_args
falsepositives:
  - Legitimate Python automation scripts that pipe output through shell commands (e.g., subprocess for system administration)
  - Package manager operations invoked from the IDE or build pipeline
  - Container entrypoint scripts that use Python as intermediate shell launcher
level: high

Known false positives:

  • Legitimate Python system administration scripts using subprocess
  • Container build processes during image creation
  • Package manager invoked via Python wrapper
  • Development workflows with shell-invocation patterns

Tuning notes: Baseline normal Python subprocess usage in your environment. Whitelist known container images and CI/CD builder agents. For Odysseus-specific deployments, exclude the odysseus container's own entrypoint scripts. Tune the ARGV pattern matching to avoid false positives from legitimate curl/wget for package downloads. Consider correlating with web API events (TL-2026-1921-SPL-001) for higher confidence.

8 more detection rules (Splunk SPL, Microsoft KQL, Sigma) · unlock with TL Creds. Sign up free to unlock the detection rules or query them with the Threadlinqs MCP server (Purple)
  • Odysseus Task Creation with Admin-Only Action Followed by Type Flip — KQL · unlock with TL Creds
  • Odysseus Python Process Shell Execution with Suspicious Arguments — KQL · unlock with TL Creds
  • Odysseus Lateral Movement and Data Exfiltration — KQL · unlock with TL Creds
  • Odysseus Task API Sequential Endpoint Access — Authorization Bypass · unlock with TL Creds
  • Odysseus Subprocess Shell Execution — Sensitive File and Command Access · unlock with TL Creds
  • Odysseus Persistent Shell Task Execution and SSH Lateral Movement · unlock with TL Creds
  • Odysseus Task API Authorization Smuggling — Sequential Create-Flip-Execute Pattern · unlock with TL Creds
  • Odysseus Scheduled Task Persistence and SSH Lateral Movement · unlock with TL Creds

Indicators of compromise

As of 2026-08-06, Threadlinqs Intelligence tracks 6 indicators of compromise for TL-2026-1921.

6 indicators of compromise · unlock with TL Creds. Compare plans

Technical analysis

Fixed in version 1.0.2. Exposed data includes user password hashes, TOTP secrets, stored provider API keys, database contents, and SSH keys to remote machines.

Odysseus is a multi-user, privacy-focused self-hosted AI workspace created by PewDiePie (Felix Kjellberg, handle archdaemon) providing chat with LLMs, autonomous agents, tools, model serving, email, and research capabilities. The project launched on May 31, 2026, rapidly accumulating over 62,000 GitHub stars and 300+ contributors. The threat model (THREAT_MODEL.md) defines a clear role-based access control (RBAC) boundary: shell and Python execution, file read/write, email, MCP tools, and vault access are admin-only capabilities. A scheduled tasks mechanism allows users to create jobs that run on schedules, webhooks, or manual triggers, with built-in actions including run_local, run_script, and ssh_command that execute shell commands on the host via Python's subprocess.run(script, shell=True).

Manifold Security discovered a critical authorization bypass allowing any authenticated non-admin user to execute arbitrary OS commands. The core defect was a state desynchronization across three code paths: the create gate in routes/task_routes.py checked whether the incoming request's task_type was "action" and the action was admin-only; the update gate checked whether the incoming request included an action field; and the scheduler dispatch in src/task_scheduler.py checked the stored task.task_type. No single code path asked whether the task as it stood would run a shell command — each evaluated only the fields the request happened to contain. By splitting the attack into two API calls, an attacker could assemble a privileged job from two individually unprivileged requests, with neither gate recognizing the full picture.

The first request creates a task with task_type set to "research" (a benign type bypassing the create gate) but attaches the admin-only run_local action. The create gate saw task_type="research" — not "action" — and never inspected the action field, persisting it to the database regardless. The second request updates only the task_type to "action" via a PUT endpoint. The update gate defaulted the action field to None in the request body, so the guard if req.action is not None evaluated as false and the admin privilege check never ran. With both modifications complete, the task carried task_type="action" (from the update) and action="run_local" (persisted from creation). Triggering the task via POST /api/tasks/{id}/run caused the scheduler to dispatch based on the stored task_type, executing arbitrary shell commands via subprocess.run(script, shell=True) with the privileges of the Odysseus process (uid 1000 in Docker).

Commands executed as the odysseus user with no sandboxing, filesystem confinement, or egress filtering. The project's own THREAT_MODEL.md acknowledges the absence of shell/filesystem sandboxing as a known gap. The process holds all sensitive application data: user password hashes (bcrypt) and TOTP secrets, stored provider AI API keys, the full application database, and SSH keys that Odysseus uses to access remote managed machines. On any instance with self-service signup enabled or more than one user, compromising a single non-admin account becomes a full foothold — attackers gain API keys to spend, a mailbox to send from, SSH keys to managed machines, and a scheduler mechanism for persistence. There is no evidence the issue was exploited before the fix.

The project consolidated five independent reports of the same defect. Manifold Security (Francisco Rosales) disclosed with a full Docker Compose PoC on June 8, 2026. The fix merged on July 5, 2026 in PR #5235 (commit 2826dcfc), binding the authorization decision to the task's effective final (task_type, action) pair, revalidating persisted state at every boundary (resume, manual-run, webhook, and scheduler dispatch), and pausing existing privileged rows owned by non-admin users rather than executing them. Odysseus has had additional security findings including CVE-2026-70619 (broken access control in embeddings endpoint, CVSS 8.8), CVE-2026-70620 (SSRF in embeddings endpoint, CVSS 6.8), and multiple prompt injection vulnerabilities (fixed in PR #3383 via untrusted context wrapping). These findings underscore the security challenges of AI-assisted development at scale — the project was reportedly "vibe coded" (AI-generated from scratch), contributing to authorization gaps across the codebase.

MITRE ATT&CK techniques by tactic:

Collection

T1005 Data from Local System

Lateral Movement

T1021 Remote Services

Exfiltration

T1048 Exfiltration Over Alternative Protocol

Execution

T1053 Scheduled Task/Job; T1059 Command and Scripting Interpreter

Discovery

T1057 Process Discovery; T1082 System Information Discovery

Command and Control

T1071 Application Layer Protocol

Initial Access

T1078 Valid Accounts

Privilege Escalation

T1548 Abuse Elevation Control Mechanism

Credential Access

T1552 Unsecured Credentials; T1555 Credentials from Password Stores

Remediation

Patches

  • Upgrade to Odysseus version 1.0.2 or cherry-pick commit 2826dcfc on the dev branch

Immediate actions

  • Upgrade Odysseus to version 1.0.2 (commit 2826dcfc) immediately on all instances
  • Audit all existing tasks across all user accounts for hidden admin-only actions
  • Rotate all credentials stored in Odysseus: provider API keys, SSH keys, user passwords, TOTP secrets

Workarounds

  • Disable self-service user registration if not operationally required
  • Review all non-admin user accounts and remove unused or suspicious accounts
  • Restrict outbound network access from Odysseus containers to only required LLM API endpoints
  • Monitor API logs for the task-type-flip pattern: POST /api/tasks followed by PUT /api/tasks/{id} changing task_type from non-admin to action, followed by POST /api/tasks/{id}/run

Longer-term hardening

  • Implement authorization revalidation at every execution boundary (create, update, resume, manual-run, webhook, scheduler dispatch)
  • Deploy runtime monitoring to detect application processes spawning child shell processes (/bin/sh) as a detection signal
  • Add network egress filtering and filesystem sandboxing for Odysseus containers (e.g., seccomp, AppArmor, read-only root filesystem)
  • Maintain a coordinated vulnerability disclosure (CVD) program with GitHub Security Advisories enabled
  • Conduct a full security audit of the API authorization middleware for state-desynchronization patterns

Affected and fixed versions

  • Odysseus — AI Workspace
    Vulnerable versions: through 1.0.1; before commit 2826dcfc
    Fixed in: 1.0.2 (commit 2826dcfc)

Timeline

  • Odysseus AI Workspace publicly announced by PewDiePie (Felix Kjellberg) via YouTube launch video; repository reaches 1.5M+ views on launch video in 24 hours and 62K+ GitHub stars
  • Multiple prompt injection vulnerabilities reported in Odysseus (PR #3383); GitHub Security Advisories not yet enabled on the repository; community contributors begin source code audits
  • Manifold Security (Francisco Rosales) independently discovers the RCE authorization bypass vulnerability and reports it with a full Docker Compose reproduction PoC
  • Within 8 days of initial disclosure, four other independent researchers reported the same RCE authorization bypass vulnerability to Odysseus maintainers, confirming the bug's discoverability
  • Additional vulnerabilities disclosed: CVE-2026-70619 (broken access control in /api/embeddings/endpoint, CVSS 8.8) and CVE-2026-70620 (SSRF in same endpoint, CVSS 6.8); both assigned by VulnCheck
  • Odysseus version 1.0.2 released containing the RCE fix (commit 9844a2f9, including fix commit 2826dcfc); default branch updated with the security patch
  • Fix merged in PR #5235 (commit 2826dcfc): authorization decision bound to the task's effective final (task_type, action) pair; persistent state revalidated at resume, manual-run, webhook, and scheduler dispatch boundaries; existing privileged non-admin rows paused
  • GitHub Security Advisory GHSA-xwhc-f36c-v5vm drafted and assessed as Critical (CVSS 9.9); pending CVE assignment from GitHub Security Lab
  • Manifold Security publishes full technical disclosure blog post detailing the exploit chain, authorization bypass mechanism, impact analysis, and fix guidance
  • The Hacker News publishes ThreatsDay coverage bringing the vulnerability to broad public attention alongside other security news

Sources cited

Previous and next in vulnerability

Drafted by the Threadlinqs research pipeline from 12 sources; published ; last updated —.

Threadlinqs Intelligence — Real-Time Threat Detection Platform

// threat_feed
$ sort --newest
Showing all threats

Live intelligence console

Threat level

Latest Threats