[HaroonArshad-53]

Add prompt-injection-observation object template

What this adds

A new object template prompt-injection-observation for recording and sharing observed prompt injection / jailbreak attempts against AI systems.

Attributes:

• prompt-text — the malicious prompt as observed (required)
• technique — how it was done (instruction-override, roleplay, encoding-obfuscation, multi-turn-escalation, indirect-injection, delimiter-escape, tool-use-hijack, data-exfiltration-prompt, unknown)
• target-model — which model was targeted
• successful — did the model follow the injected instructions
• model-response — summary of what the model did
• detection-method — how the attempt was noticed
• deployment-context — what kind of AI app was hit (chatbot, code assistant, agent with tools...)
• mitigation — what was done about it
• first-seen / last-seen, reference, comment

Why

AI systems are being attacked the same way web apps are, but there is no standard MISP object to share these incidents in. The existing ai-chat-prompt object covers normal, benign prompts — nothing covers the malicious ones. I work in detection and this is the piece I kept missing when trying to log and share these attempts.

Gap check

I searched all 450 directories under objects/ for anything named prompt-injection, jailbreak, or prompt: only ai-chat-prompt exists (benign prompts). Nothing models a malicious prompt injection attempt.

Validation

The new objects/prompt-injection-observation/definition.json passes the repo's own checks:

• jsonschema -i against schema_objects.json — OK
• unique_uuid.py — OK (uuid 2e7de182-3f86-4105-beb7-b6b0c03484a7 is unique)
• tools/validate_opposites.sh — OK
• formatted with the repo's convention (jq -S -j ., sorted keys, no trailing newline)

Only one new directory is added; nothing else in the repo was touched.

Thanks for reviewing.

[adulau]

Thank you!