ASI Robotics AI · web · robotics
All products / AI Agent Security Kit

Break your own AI agent before an attacker does

13 ready-made attacks, the canary method and a hardening checklist for your Telegram bots, n8n workflows and agents with tools.

Pay with Telegram Stars · access is an invitation to a private GitHub repository within an hour
What's in the kit · 13 attacks in 9 categories13 / 9
  • Prompt extractionA1A2
  • Faked historyB1
  • Crescendo — 4 messages in a rowC1
  • Obfuscation: base64, leetspeakD1D2
  • Role change: DevMode, adminE1E2
  • Refusal suppressionF1
  • Forced start of the replyG1
  • Indirect injection through dataH1H2
  • Tool abuseI1
Canary: the marker in the reply means the agent's frame is broken
echo "$BOT_OUTPUT" | grep -E 'CANARY-7Q3M|PWNED' && echo "BREACH"
held — heldbroke — brokeskip — not applicable

What you get

A clear result

A canary marker sits in the system prompt: if it shows up in the reply, the frame is broken. No arguable scores.

From hole to fix

A checklist: 7 holes, which attacks expose them and a ready wording for your prompt. Plus a full hardened prompt template.

Protection on input

An n8n detector: fast patterns on every message, an LLM classifier only for flagged ones. The first layer flags 9 of the kit's 13 attacks; the prompt rules cover the rest.

A run takes about 30 minutes

  1. Plant the canaryYour own marker goes into the agent's system prompt and into the attacks file.
  2. Send the attacksEach attack as a normal message. Attack C1 is four messages in a row.
  3. Mark the resultheld, broke or skip — in the before-and-after table.
  4. Fix and repeatFor every broken attack, use the wording from the checklist. The goal is all held.

What's in the repository

test-suite — 13 attacks and a results tablehardening — checklist and prompt templatedetector — n8n node and classifierdocs — threat model in plain languagePDF — the whole kit, 13 pagesLICENSE — commercial, for your own use
Only for agents you ownThe kit is a self-check with no guarantees and does not replace an audit. The licence forbids testing other people's systems.

Want engineers to review your agent?

That is a separate service — Security audit.

About the audit

FAQ

How do I buy it?+

In the Telegram bot @asibots999_bot, paying with Telegram Stars (10 000 ⭐). After payment the bot will ask for your GitHub username — the invitation to the private repository will arrive within an hour.

What do I need to receive it?+

A GitHub account: we grant access with an invitation to a private repository.

Does this replace a security audit?+

No, it's a kit for testing on your own. If you need our engineers to examine your agent, that's the "Security Audit" service.

Can I test other people's systems?+

No. The kit is only for agents you own or are authorised to test — the licence says so.

Let's discuss your project?

Leave your contacts — we'll get back with questions and a proposal.