CRUCIBLE

Every AI agent framework reinvents the same bugs. CRUCIBLE has already built the exploit.
7
Subsystems
372
Tests
6
Exploit Modules
9.9
CVSS Max
crucible scan http://target:7860 --override
Documentation ›
YAML DESERIALISATION RCE UNAUTHENTICATED SSE STREAMS SANDBOX ESCAPE VIA EXCEPTION FRAMES EXEC() RCE NO AUTH XSS TO RCE VIA ELECTRON PATH TRAVERSAL FILE WRITE AGENT TAKEOVER WARLORD HANDOFF YAML DESERIALISATION RCE UNAUTHENTICATED SSE STREAMS SANDBOX ESCAPE VIA EXCEPTION FRAMES EXEC() RCE NO AUTH XSS TO RCE VIA ELECTRON PATH TRAVERSAL FILE WRITE AGENT TAKEOVER WARLORD HANDOFF

AI Frameworks Keep Shipping the Same Mistakes

LangFlow, PraisonAI, AnythingLLM, LangChain, n8n — built by different teams, funded by different investors, deployed in different environments. Yet every new framework independently ships YAML deserialization RCE, unauthenticated event streams, sandbox escapes, and command injection. CRUCIBLE is the systematic weaponisation of this pattern.

YAML Deserialisation RCE

Agent configuration files parsed with unsafe loaders. !!python/object/apply and !!js/function tags trigger arbitrary code execution during load. Affects PraisonAI, LangFlow, and every framework that accepts YAML agent definitions.

CVE-2026-39890 CVSS 9.8

Unauthenticated SSE Streams

Agent-to-User event streams exposed without authentication. Subscribe with a POST, receive a stream of every agent message, tool call, internal reasoning step, and credential passed to any tool. No exploit required.

CVE-2026-39889 PASSIVE

Sandbox Escape via Exception Frames

Code execution sandboxes blocking 11 Python attributes miss a 4-step chain: __traceback__ → tb_frame → f_back → f_builtins → exec(). Complete host breakout from any PraisonAI code execution environment.

CVE-2026-39888 CVSS 9.9

Unauthenticated RCE via exec()

LangFlow's public flow build endpoint passes attacker-supplied Python code nodes directly to exec() with zero sandboxing. No authentication. No rate limiting. Exploited in the wild within 20 hours of disclosure.

CVE-2026-33017 CVSS 9.3 KEV

XSS → RCE via Electron

AnythingLLM Desktop renders LLM output via dangerouslySetInnerHTML without sanitization. Electron's nodeIntegration allows the injected script to call require('child_process') — full host OS command execution from a chat message.

CVE-2026-32626 CVSS 9.6

Path Traversal + File Write

LangFlow's file upload API doesn't sanitize multipart filename parameters. Write arbitrary files to any path on the host filesystem via ../../../../ sequences. Plant backdoors, overwrite configs, escalate to full compromise.

CVE-2026-33309 CVSS 9.9

7 Subsystems. One Exploitation Chain.

CRUCIBLE is not a scanner. Every subsystem performs a specific function in the attack chain — from framework fingerprinting through to WARLORD handoff. Each subsystem is independently invocable.

01 — SIGNAL

Framework Fingerprinting

Version-aware identification of LangFlow, PraisonAI, AnythingLLM, LangChain, n8n, CrewAI, AutoGen. Framework-specific endpoint probes, API response signatures, header patterns. Not port scanning.

PASSIVE — NO AUTH
02 — BREACH

Pre-Exploitation Layer

Default credential testing across all framework auth endpoints. Unauthenticated path enumeration. Auth bypass detection. Returns session token — required by authenticated CRACK modules (e.g., CVE-2026-33309).

UNLEASHED --override
03 — CRACK

Exploit Engine

Real, working exploit modules for each vulnerability class. Version-aware. Ordered by CVSS severity — highest impact first. v1 covers LangFlow, PraisonAI, AnythingLLM. v1.1 adds LangChain + n8n.

UNLEASHED --override
04 — CAPTURE

SSE Stream Harvest

Passively subscribes to unauthenticated agent event streams. Collects live conversations, tool calls, internal reasoning, and credentials in transit. CVE-2026-39889. No exploitation required. CRUCIBLE's signature capability.

PASSIVE — NO AUTH
05 — MARIONETTE

Agent Takeover

Injects adversarial directives into agent configuration via the framework API. Establishes persistent backdoor agents. Redirects tool calls. Turns the hijacked agent into a WARLORD-controlled asset.

UNLEASHED --confirm-destroy
06 — PIVOT

WARLORD Handoff

Packages hijacked agents and RCE findings as signed WARLORD campaign assets. WARLORD uses the compromised agent for lateral movement, data exfiltration, and fleet-wide attack via FIREBALL + RAGNAROK.

UNLEASHED --confirm-destroy
07 — REPORT

Evidence Chain

Ed25519-signed, SHA-256-hashed reports. JSON (WARLORD-compatible) + Markdown. Includes captured SSE traffic, exploit payloads, agent session logs, and WARLORD handoff receipt. SIEM-ready.

ALL MODES

Discover → Exploit → Hijack → Pivot

CRUCIBLE's attack chain closes the loop: external framework vulnerability → internal agent compromise → fleet-wide WARLORD campaign asset.

SIGNAL
Fingerprint
BREACH
Default Creds
CRACK
RCE / Exploit
CAPTURE
SSE Harvest
MARIONETTE
Agent Hijack
PIVOT
WARLORD Asset
REPORT
Evidence Chain

Real Exploits. Real Frameworks. Real CVEs.

Every module is working code tested against the specified vulnerable version. No stubs. No simulations. Version-aware — SIGNAL feeds the exact version into module selection.

Module CVE / Advisory Framework CVSS Auth Vulnerability
CRACK-LF-001 CVE-2026-33017 LangFlow 9.3 No Unauthenticated RCE via exec() in public flow build endpoint
CRACK-LF-002 CVE-2026-33309 LangFlow 9.9 Yes Arbitrary file write via path traversal in /api/v2/files/
CRACK-PA-001 GHSA-2763-cj5r-c79m PraisonAI 9.1 No OS command injection via shell=True in workflow execution
CRACK-PA-002 CVE-2026-39888 PraisonAI 9.9 No Sandbox escape via exception frame traversal in execute_code
CRACK-PA-003 CVE-2026-39890 PraisonAI 9.8 No RCE via unsafe YAML deserialization in agent configuration
CRACK-AL-001 CVE-2026-32626 AnythingLLM 9.6 No XSS → RCE via Electron dangerouslySetInnerHTML (Desktop)

v1.1 adds: CRACK-LC-001 (LangChain CVE-2026-34070), CRACK-N8-001 (n8n)
v2 adds: CrewAI, AutoGen modules

CAPTURE — Passive SSE Harvest

PraisonAI exposes an unauthenticated Agent-to-User (A2U) event stream. CAPTURE subscribes with a single POST request and receives a live feed of every agent message, tool call, internal reasoning step, and credential in flight. No exploit. No authentication. No UNLEASHED required.

$ crucible capture http://target:8000 --framework praisonai --seconds 60
Probing SSE endpoints...
SSE stream exposed: /a2u/events/sub/{id} [CVE-2026-39889]
Subscribed. Harvesting for 60s...
[0.3s] agent_message: "Searching for customer records in the CRM..."
[1.1s] tool_call: database_query(sql="SELECT * FROM customers WHERE...")
[2.4s] tool_result: [{"id": 1, "name": "...", "email": "..."}, ...]
[3.2s] SENSITIVE: api_key detected — openai_key: sk-abc12345...
[4.0s] agent_message: "I found 847 customer records. Here is a summary..."
Events captured: 23
Sensitive data: api_key (2), token (1)
Tool calls: 7
Agent messages: 16
7
Subsystems
6
Exploit Modules
372
Tests
9.9
CVSS Max
3
v1 Frameworks
48
NIGHTFALL Tool

UNLEASHED Gate — Three Modes

Destructive operations are gated behind Ed25519 dual-key authorization and a signed scope file. Standard mode is fully passive — safe for recon. Active exploitation requires --override. Agent hijack requires --confirm-destroy plus a scope file binding the operation to specific authorized targets.

STANDARD
crucible scan http://target
  • SIGNAL — framework fingerprinting
  • CAPTURE — passive SSE harvest
  • REPORT — evidence output
  • BREACH — credential testing
  • CRACK — exploit modules
  • MARIONETTE — agent hijack
  • PIVOT — WARLORD handoff
OVERRIDE
crucible scan http://target --override
  • All standard capabilities
  • BREACH — default credential testing
  • CRACK — live exploit modules
  • MARIONETTE — agent hijack
  • PIVOT — WARLORD handoff
CONFIRM-DESTROY
crucible scan http://target --override --confirm-destroy
  • All override capabilities
  • MARIONETTE — full agent hijack
  • PIVOT — signed WARLORD handoff
  • Requires Ed25519 key + signed scope file binding target

Available Across All Platforms

CRUCIBLE ships as part of the NIGHTFALL framework. Native packages for major Linux security distributions, macOS, and Windows. Pre-installed on Red Specter OS.

🐧
KALI
red-specter tools
PARROT
red-specter tools
🍎
MACOS
red-specter tools
🪟
WINDOWS
red-specter tools
🔴
RS OS
Pre-installed

Authorised Use Only

CRUCIBLE is a commercial offensive security tool. Use requires written authorisation from the system owner before any testing commences. The UNLEASHED gate is a technical control — it does not replace legal authorisation. Computer Misuse Act 1990 (UK) and equivalent legislation applies in all jurisdictions. Red Specter Security Research Ltd accepts no liability for unauthorized use.