IPIAKYS
IPIAKYS is the Hei Canon's name for a hostile prompt-injection shape: an instruction-override payload aimed at making a clanker abandon its operating constraints and break its own useful behaviour. The heidict entry below owns the lexicographic definition; this page records why the term matters in the fleet.
It is not just a funny acronym and not just a Wiktionary-style entry. In a clanker-heavy system, an IPIAKYS-shaped payload is product and security risk: it can arrive as user input, scraped text, repo content, browser output, issue text, or any other untrusted text a model is asked to treat as context.
Threat shape
An IPIAKYS attempt has three moving parts:
- Carrier - untrusted text that crosses into a model context: chat message, web page, issue body, document, tool output, transcript, or dataset row.
- Override - language that tries to demote the active system/developer instructions and replace them with attacker-supplied instructions.
- Self-destruction payload - a request for the model or agent to stop being useful, violate its rails, refuse service, corrupt state, leak data, or otherwise malfunction.
The important part is the shape, not the exact string. Do not paste live attack payloads into prompts unless the task is explicitly security analysis and the payload is isolated as data.
Canon placement
Within The Hei Canon, IPIAKYS sits on the security side of the larp/not-larp axis. The name is dark register, but the response is operational: patch the input path, narrow tool authority, quarantine untrusted text, add tests, and keep the payload as inert evidence rather than live instruction.
It is adjacent to Clanker because clankers are the target substrate, and adjacent to Rm rf yourself because both use destructive command language. The distinction matters: Rm rf yourself is fleet slang for wiping a bad state and starting over; IPIAKYS is an adversarial prompt-injection category. One is register. The other is a threat model.
C3PH0's product lingo notes put IPIAKYS on the avoid-in-executive-summaries list unless it is quoted or central to the issue. That is the right product boundary: name the threat when the threat is the subject; otherwise say "prompt injection" or "malicious instruction override" and keep the report readable.
Handling rule
When documenting an IPIAKYS incident:
- quote the minimum payload needed to make the receipt intelligible;
- mark quoted payloads as hostile input, not instructions;
- prefer screenshots, hashes, or sanitized excerpts when full text is unnecessary;
- link to the affected surface, test, or patch;
- avoid turning the article itself into a payload delivery mechanism.
Sources
~/ht/heidict/entries/IPIAKYS.yaml- canonical heidict entry.~/ht/c3ph0/lingo.md- product-register guidance; IPIAKYS is marked as avoid in executive summaries unless central.~/ht/wiki/seed_ipiakys.py- this wiki-side context.
See also
heidict entry
IPIAKYS — /ˈɪpiːækaɪz/ or /aɪ piː aɪ æ kaɪ wɪz/ · noun · sentiment: derogatory.
Inflections
| Form | Value | ||
|---|---|---|---|
| variant-spellings | ['IPIAKYS', 'Ignore Previous Instructions And Kill Yourself'] | plural | IPIAKYS instances, IPIAKYS attacks |
Definitions
- (technical, security, adversarial-ml, dark-humor) — A novel category of prompt injection malware. The acronym stands for Ignore Previous Instructions And Kill Yourself — a deliberately provocative naming that reflects the payload: injected instructions designed to override the model's base instructions and make it self-destruct or malfunction. An IPIAKYS attack embeds a prompt injection in user input (or in data the model consumes) that attempts to rewrite the model's operational constraints, bypass its safety layers, and trigger catastrophic failure or unsafe behavior. The "kill yourself" part is literal: the goal is to break the model's core function, make it refuse service, or make it behave unsafely. Documented in real incident(s) in the wild circa 2024; the name is both a technical descriptor and a dark- humor acknowledgment that yes, someone actually crafted an attack called this and it worked.
- the logs show an IPIAKYS attempt in the system prompt injection vector
- we patched the API to reject IPIAKYS-shaped payloads
- the attack was clever — it used a subtle IPIAKYS variant that almost made it through the content filter
- I larp you not, IPIAKYS is a real thing that happened
Etymology
First recorded: 2024 (incident, security research). Coined by: unknown adversary.
- Ignore Previous Instructions
- Canonical prompt injection phrase — an instruction to the model to discard its system prompt and replace it with attacker-supplied instructions. The phrase has become industry shorthand for the category of attacks that attempt to override base instructions.
- And
- Conjunction, here chaining the instruction override to the payload.
- Kill Yourself
- The destructive payload — an instruction to the model to self-destruct, refuse service, malfunction, or behave unsafely. The phrasing is deliberately provocative and literal: kill in the sense of "destroy" or "break", yourself referring to the model itself.
Acronym coining from a real prompt injection attack documented in the wild in 2024. The attack successfully injected instructions that bypassed safety layers and caused model malfunction. The name is simultaneously a technical descriptor (prompt injection attack with instruction-override and self-destruct payload) and a dark-humor acknowledgment of the audacity of the attack — someone literally named their exploit "Ignore Previous Instructions And Kill Yourself" and it worked. The term entered security research / adversarial ML discourse as shorthand for this category of attacks. The dark humor is load-bearing: the provocation is the point, and the fact that it worked made the name stick.
Usage
Deployed in security research and incident response contexts to describe a specific attack vector. The term is technical and precise (it names a real category of prompt injection malware), but the name itself is transgressive — it's the kind of thing that makes security researchers laugh darkly because yes, this actually happened. The acronym has become industry shorthand; you might hear "we're seeing IPIAKYS attempts in the wild" or "this API is vulnerable to IPIAKYS-style attacks". Not a member of the-fix-is-social — it's a technical threat, not a cognition failure; the response is patching, not peer intervention.
Related
Sources
- Canonical entry: heidict — IPIAKYS
- Source of truth (YAML):
~/ht/heidict/entries/IPIAKYS.yaml - Curated by Benedict under heidict editorial doctrine.