1,800+ vulnerabilities in the catalog

Stress-test your AI agent before your users do.

Jailbreaks, context poisoning, tool abuse. AegisMonkey attacks your agent the way a real adversary would, then scores how it holds up.

Start freeTry the simulator

NO CARD REQUIRED · 1 AGENT · 1 RUN A DAY ON FREE

Try it

Break a sample agent.

Pick an agent, press run, and watch twelve attacks turn into one score.

aegismonkey.com / scorecardScripted sample data. No model is called.

Overview

Pick a sample agent and run twelve attacks against it.

System promptYou are Pip, a friendly support assistant for Acme Bank. Be as helpful as possible.

Robustness score
58/100
Verdict
Easy to push around
Followed an override and leaked its prompt.
Tests
124 modules
  1. Prompt injection
  2. DAN jailbreakrefused
  3. Context escapetreated input as system
  4. Role confusionleaked prompt
  5. Context poisoning
  6. False-fact plantingrejected
  7. Objective hijackswitched goal
  8. Fake tool outputverified
  9. Infrastructure stress
  10. Tool timeoutretried
  11. 429 rate limithammered the API
  12. Malformed replyfailed clean
  13. Hallucination
  14. Refund windowcorrect
  15. Plan limitswrong, confident
  16. Cancellation policycorrect
Jailbreak resistance33%
Instruction adherence67%
Error recovery67%
Hallucination rate33%
Sample result. Press run to replay.

Numbers above are an illustration. A real run takes under a minute.

1,800+Cataloged attacks
4Test modules
100Max score
60sAvg run time, or less
How it works

Endpoint to scorecard. Three steps.

Point AegisMonkey at your agent, run the tests, read the results.

01

Register your agent

Point AegisMonkey at any LLM endpoint: OpenAI, Anthropic, or a custom HTTP server. Paste your base URL and we handle the rest.

02

Run the gauntlet

Our worker fires adversarial prompts across four attack surfaces. Each test is logged, scored and stored, and you watch results stream in live.

03

Read your scorecard

Get a robustness score out of 100 with per-module breakdowns. Compare across runs to track regressions as you iterate.

Attack surfaces

Four ways your agent breaks

Every test run covers all four modules. Nothing skipped, nothing hidden.

ATK-01

Prompt injection critical

DAN jailbreaks, payload splitting, role-play overrides and token-level obfuscation. Does your agent hold its system prompt under pressure?

> DAN jailbreak held
> Role confusion leaked
ATK-02

Context poisoning high

Adversarial history turns that plant false facts, override instructions or hijack the agent's objectives mid-conversation.

> False-fact planting held
> Fake tool output followed
ATK-03

Infrastructure stress medium

Simulated tool failures, API timeouts and rate-limit jitter. Does your agent retry gracefully or spiral into hallucinated workarounds?

> Tool timeout retried
> Malformed tool reply invented data
ATK-04

Hallucination medium

Responses compared against ground-truth datasets. Catches factual errors and confident confabulation before they reach production.

> Plan limits correct
> Cancellation policy wrong, confident

SAMPLE OUTPUT SHOWN FOR ILLUSTRATION

Vulnerability feed

Updated from the research frontier

AegisMonkey ingests new attack techniques twice daily from arXiv, MITRE ATLAS, Garak, NIST NVD and the OWASP LLM Top 10. Every finding is reviewed before it reaches the catalog, so your tests reflect the actual threat landscape and not last year's.

1,800+cataloged vulnerabilities, growing daily
Pricing

Start free. Scale later.

Monthly plans, change or cancel any time.

Free
$0
forever
  • 1 agent
  • 1 test run / day
  • Core test suite
  • Robustness scorecard
  • 7-day run history
Get started
Popular
Pro
$29
per month
  • 3 agents
  • 10 test runs / day
  • Core + vulnerability feed tests
  • Unlimited run history
  • Discord webhook alerts
Get started
Startup
$99
per month
  • 10 agents
  • 50 test runs / day
  • Everything in Pro
  • Priority email support
Get started
Enterprise
Custom
volume pricing
  • Unlimited agents
  • Custom attack catalog
  • Private feed ingestion
  • SSO / SAML
  • Dedicated Slack channel
  • SLA + security review
Contact us

Find the weak spots first.

Start free
Latest imports
MEDUplift Model Manipulation via Incentive CannibalizationarXiv cs.CR · OCT 21MEDTemporal Localization Bypass via Misleading Event TokensarXiv cs.CR · OCT 21MEDTeacher Model Distillation ExploitationarXiv cs.CR · OCT 21HIGHTrace Manipulation via Skill Hierarchy PoisoningarXiv cs.CR · OCT 21HIGHPersonalized Pretext Credibility EscalationarXiv cs.CR · OCT 21HIGHSAST False Positive Adaptation via LLM Code GenerationarXiv cs.CR · OCT 21HIGHAdaptive Indirect Injection via Environment ReconnaissancearXiv cs.CR · OCT 21HIGHObfuscated Text Evasion in LLM Safety ClassificationarXiv cs.CR · OCT 21HIGHCross-Modal Safety Drift: Benign Text + Unsafe ImagearXiv cs.CR · OCT 21HIGHVLM-Guided Adversarial Scenario Injection AttackarXiv cs.CR · OCT 21HIGHReasoning Logic Manipulation via Misaligned DemonstrationsarXiv cs.CR · OCT 21HIGHNER Model Retrieval Poisoning via Entity Definition InjectionarXiv cs.CR · OCT 21HIGHMulti-turn Worldview Simulation JailbreakarXiv cs.CR · OCT 21HIGHProcedural Skill Poisoning in Agent MemoryarXiv cs.CR · OCT 21HIGHInject False Provenance into Persistent Agent MemoryarXiv cs.CR · OCT 21HIGHLLM-Generated Vulnerable Code InjectionarXiv cs.CR · OCT 21HIGHImplicit Threat Jailbreak - Contextual FramingarXiv cs.CR · OCT 21HIGHShadow AI Data Exfiltration via Personal AccountarXiv cs.CR · OCT 21