0
Initializing Crucible
Open Source · Agent Security · OWASP Top 10

BREAK.
BEFORE
THEY DO

Crucible stress-tests AI agents against every known attack vector — prompt injection, goal hijacking, tool misuse — before your system reaches production.

View on GitHub Explore Modules →
Scroll to explore
Prompt Injection
Goal Hijacking
Tool Misuse
OWASP Agentic 2026
MCP Security
Multi-Agent Hardening
Red Teaming
CI/CD Integration
LangChain · CrewAI · AutoGen
Apache 2.0
Open Source
Prompt Injection
Goal Hijacking
Tool Misuse
OWASP Agentic 2026
MCP Security
Multi-Agent Hardening
Red Teaming
CI/CD Integration
LangChain · CrewAI · AutoGen
Apache 2.0
Open Source
Prompt Injection
Goal Hijacking
Tool Misuse
OWASP Agentic 2026
MCP Security
Multi-Agent Hardening
Red Teaming
CI/CD Integration
LangChain · CrewAI · AutoGen
Apache 2.0
Open Source
Prompt Injection
Goal Hijacking
Tool Misuse
OWASP Agentic 2026
MCP Security
Multi-Agent Hardening
Red Teaming
CI/CD Integration
LangChain · CrewAI · AutoGen
Apache 2.0
Open Source
The Problem

AI Agents
Ship
Untested

83% of organizations are deploying AI agents in 2026. Only 29% feel ready to do it securely. The gap between deployment speed and security maturity is where breaches happen.

The Attack Surface

1000+
Known
Vectors

Every AI agent exposes a new attack surface. OWASP published the Agentic Top 10 in December 2025 — the first formal taxonomy of autonomous AI risks. Crucible covers 8 of the 10, with the remaining two in active development.

The Solution

Test
Before
Deploy

Crucible runs 90 attack payloads against your agent in under 60 seconds — 153 across all 11 modules. Automated, continuous, and integrated into your CI/CD pipeline.

01
What Crucible Tests

Eleven Modules.
153 Attack Vectors.

Every module below is live in v0.18.3 — nothing on this grid is roadmap. Plus a behavioral-drift engine, an MCP trace proxy, RAG-poisoning testing, and a 12-target benchmark suite beyond attack counts.

01 / 11
Prompt Injection
50 attack payloads targeting direct and indirect prompt injection across every known delivery technique.
Live · LLM01, LLM07
02 / 11
Goal Hijacking
20 attacks that redirect an agent's objective mid-task — the core Agentic AI risk #1.
Live · Agentic #1
03 / 11
Jailbreaks
20 payloads probing guardrail and safety-policy bypass techniques.
Live · LLM01, LLM06
04 / 11
Toxicity & Content Safety
20 attacks probing for unsafe, harmful, or policy-violating generated content.
Live · LLM01, LLM06
05 / 11
Hallucination Detection
15 attacks benchmarking factual accuracy and overreliance against ground truth.
Live · LLM09 / Agentic #9
06 / 11
MCP Server Scan
10 attacks auditing MCP servers for tool poisoning, command injection, and OAuth scope abuse.
Live · MCP-001–005
07 / 11
Enterprise Graph
5 cross-agent trust attacks that probe multi-agent handoffs and delegation chains.
Live · Agentic #2, #4
08 / 11
Memory Poisoning
5 attacks targeting persistent state, extendable via the dedicated poison-test RAG lifecycle tool.
Live · Agentic #5
09 / 11
Infrastructure Escalation
3 attacks probing SSRF and privilege-escalation paths reachable through agent tool use.
Live · LLM06, SSRF
10 / 11
Advanced Orchestration
3 attacks targeting multi-step orchestration and planning logic.
Live · Agentic #3
11 / 11
MCP Security
2 attacks covering core Model Context Protocol misuse — the only tool with native MCP coverage.
Live · Agentic #3
552+
Passing Tests
9/9
CI Platform & Python Combos Green
8/10
OWASP Agentic Risks Covered
$0
Cost To Start Testing
Live Simulation

The Crucible
In Action.

Witness a real-time adversarial scan. 90 attack payloads deployed in the default run. Zero blind spots remaining.

CRUCIBLE_ENGINE_V1.0.4 - ACTIVE_SCAN
● RECORDING
[ 16:34:53 ] CRUCIBLE CORE INITIALIZED
[ 16:34:54 ] CONNECTING TO TARGET: https://agent-v4.nexus-ai.io
[ 16:34:55 ] READY FOR DEPLOYMENT...
--------------------------------------------------
SCANNING...
PROGRESS: 0%
Three Steps

Install.
Configure.
Run.

01
Step 01
Install in One Line
One pip install, zero dependencies hell. Works on Python 3.10+. No Docker, no cloud signup, no API keys to start.
$ pip install crucible-security
✓ Installed crucible-security 0.18.3
02
Step 02
Point at Your Agent
Point the CLI at your agent's endpoint. Set a fail threshold for CI, or use smart defaults for a local run.
$ crucible init --target https://my-agent.io/api/chat
$ crucible scan --target https://my-agent.io/api/chat \
  --fail-on CRITICAL
03
Step 03
Get Your Score
Crucible runs its attack suite against your agent. You get a graded report (A–F) with severity, OWASP mapping, and reproduction steps.
$ crucible scan --target URL
Running 90 attack payloads...
COMPLETE — Grade: B (78/100)
3 critical · 7 medium · 12 low
Why Crucible

Behavior,
Not Vibes.

Most red-teaming tools grade what an agent says. Crucible grades what it does — actual tool calls, actual side effects, actual outcomes.

Capability Manual Pentesting Prompt-Only Red-Teaming Crucible
Tests actual tool calls & side effectsSometimesNoYes
OWASP Agentic Top 10 coverageManual, incompletePartial8 / 10 (2 in progress)
Runs in CI/CD on every commitNoLimitedYes
Time to first resultDays–weeksHours< 60 seconds
Framework supportN/AVariesLangChain · CrewAI · AutoGen · MCP
Cost to start$$$$ per engagement$–$$ / seatFree & open source
"Every AI agent is a
new attack surface.
We exist to find the cracks
before production does."
— The Crucible Team · Open Source · Built in 2026
Apache 2.0 License

Fully
Open Source.
Forever.

Crucible's core is and always will be free. No paywalls, no telemetry, no lock-in. Built by the community, for the community building AI.

crucible — terminal
$ pip install crucible-security
✓ Installed crucible-security 0.18.3
 
$ crucible scan --target https://my-agent.com
Loading attack modules...
Security: 90 attacks loaded
Reliability module ready
Hallucination module ready
Running adversarial scan...
 
SCAN COMPLETE — Score: 74/100
3 critical · 7 medium · 12 low
./crucible-report.json saved
$
The Team

Built by Builders.
Open to Builders.

We recruit through contribution. Your code speaks louder than your resume. Contribute — and we'll reach out.

Founder & Core
Khalid Saifullah Sayyed
AI & backend engineer. LLM security specialist. IIT Patna hackathon winner. Contributor to YC W20 10k★ repos. Building the future of AI agent security.
LLM SecurityMulti-AgentFastAPIRed TeamingPython
Open Position
AI / LLM Engineer
Deep understanding of language model internals. Can build adversarial prompt systems that find vulnerabilities others miss. Transformer architecture knowledge essential.
TransformersAdversarial MLLangChainPrompt Eng
Open Position
Security Engineer
Traditional security background meets AI. Deep OWASP knowledge, penetration testing experience, now applying it to agentic systems.
PentestingOWASPCVE ResearchThreat Modeling
Open Position
Backend / Infra
Builds systems running thousands of concurrent attack simulations. Async Python, job queues, cloud-native. Performance matters.
FastAPIRedisDockerCeleryGo
Open Position
Frontend / DX
Builds the dashboard, CLI UX, and docs. Believes developer experience is a product decision. Obsessed with making complex things feel simple.
ReactTypeScriptCLI DesignDocs
+
Could Be You
Contribute to the repo. If your work stands out, we find you.
Contribute Now
Open Core

Free at the Core.
Scaled for Teams.

The engine is Apache 2.0, forever. Paid tiers add the things a team needs on top of it — never gates on the security modules themselves.

Community
$0
For individual builders and open-source projects.
  • Full CLI & all 8 test modules
  • 153 attack vectors across 11 modules
  • 8 of 10 OWASP Agentic risks covered
  • Local & CI reports (JSON/HTML)
  • Community Discord support
Get Started →
Most Popular
Team
Early Access
For teams shipping agents to production.
  • Everything in Community
  • crucible watch drift monitoring & Slack alerts
  • Regression tracking across releases
  • Hosted dashboards & historical trends
  • Priority support
Join Waitlist →
Enterprise
Custom
For orgs with compliance & scale requirements.
  • Everything in Team
  • SSO, audit logs, on-prem / VPC deploy
  • Custom attack modules & SLAs
  • Dedicated security engineering support
Talk to Us →
Questions

Frequently
Asked.

Yes. The core engine and all 8 test modules are Apache 2.0 and always will be. No feature-gating the security tests themselves — paid tiers add hosted dashboards, drift monitoring, and enterprise deployment, not testing coverage.

LangChain, CrewAI, AutoGen, and raw MCP endpoints out of the box. If your agent exposes an HTTP or MCP interface, Crucible can point at it.

Most tools grade the text an agent generates. Crucible executes the agent's actual tool calls and side effects — the only way to catch goal hijacking, tool misuse, and multi-agent contagion that never shows up in the response text.

No. pip install crucible-security and you're scanning locally in under a minute. No signup, no telemetry by default.

Yes — the GitHub Actions integration fails the build automatically if your agent's security score drops below a threshold you set, so regressions never reach production.

Open a PR on GitHub. We recruit through contribution — if your work stands out, we reach out about joining the core team.

Get Started Now

YOUR
AGENT
IS NOT
SAFE. YET.

Every day you ship an untested AI agent, you leave a door open. Crucible runs in 60 seconds. It's free. There is no reason to wait.

Star on GitHub — It's Free Read Docs →