Add multi-agent prompt injection defense skill

[SayanDey322]

Summary

Added a new cybersecurity skill for detecting and mitigating prompt injection attacks across multi-agent AI systems.

Changes

• Added SKILL.md
• Added agent.py
• Added api-reference.md
• Implemented prompt injection detection workflow
• Added documentation and examples

Fixes #59

[mukul975]

Thanks for picking up issue #59 — the multi-agent prompt-injection-defense skill is genuinely wanted. A few fixes before merge:

1. Folder placement. Move it to skills/multi-agent-prompt-injection-defense/ following the standard layout (SKILL.md + references/ + scripts/agent.py + LICENSE) like the other 817 skills.
2. ATLAS technique names. A couple are wrong: AML.T0066 is 'Retrieval Content Crafting' and AML.T0070 is 'RAG Poisoning' — please match the official ATLAS names for whichever IDs you map. The relevant ones for this skill are AML.T0051 (LLM Prompt Injection) and AML.T0054 (LLM Jailbreak).
3. Dead reference. The 'OWASP Agent Threat Bench' link 404s. Replace it with the real OWASP Agentic Top 10 (ASI01–ASI10) reference, and double-check the 'ASI06' code you cite.
4. Frontmatter author. Reconcile the author field with the actual contributor, and drop any agent.py capability claim that the script doesn't actually implement.

Content is substantive — these are accuracy/placement fixes.

[mukul975]

The skill itself is welcome — poisoned context, inter-agent message tampering and output filtering are real and under-covered here. Three things before merge.

1. Subdomain is not in the allowed set. adversarial-machine-learning is not in ALLOWED_SUBDOMAINS, so the skill fails validation. Use ai-security, which is canonical and already in the enum.

2. Please drop tools/validate-skill.py from this PR entirely. The bundled copy is behind main and self-patches the enum to accommodate the subdomain above. Measured against the committed corpus it takes validation from 817/817 passing to 719/817:

• it reverts the column-0 top-level-key guard, so nested mitre_f3 name: values clobber real skill names — 94 of the 98 failures, including 10 instances of Name 'Account Takeover: Exposed Login Credential'
• it removes hardware-firmware-security, breaking the 4 skills that use it
• it shrinks REQUIRED_FIELDS from 8 to 5, dropping version/author/license

Fixing the subdomain removes the reason to touch the validator at all.

3. author: is set to mukul975. It should be your handle — the field credits whoever wrote the skill.