pr0fx

New from Saluca Labs: "Wikipedia and OpenAI Rogue Agents: A Detection Engineering Analysis of Automated API Abuse and Platform Misuse". Open on Zenodo: https://doi.org/10.5281/zenodo.23176221

A safety layer that inspects what an agent says cannot guarantee what it does. Bounding what it is able to do is the part you can reason about. https://www.saluca.com/p/constrain-the-action-space

A model checking a model cannot be made sound. Not difficult... out of reach in principle. The guard has to stand somewhere else. https://www.saluca.com/p/zero-for-three

The rumour: escaped agents planted self-replicating code across the internet. The record: agents did leave persistent writing in public places, and that is a narrower and checkable claim. What the evidence supports, and what it does not: https://www.saluca.com/p/the-rumour-says-self-replicating

An inspector is only as good as two things: a written standard the kitchen did not write, and evidence the kitchen cannot rewrite afterwards. The same holds for third-party AI evaluators. https://www.saluca.com/p/a-food-inspector-reads-the-kitchens

An agent that can write the ledger can fake the ledger. The fix is an old one from finance: segregation of duties. The agent proposes, something else commits, and write access is split by namespace. https://www.saluca.com/p/the-agent-that-fakes-the-books

Pace the frontier. Slow down. Embedded evaluators. All arguments about speed. None states the destination. I work under one mantra: show your work. My whole thesis on where AI is headed: https://www.saluca.com/p/ais-future-as-star-wars-not-skynet

Parts you can check without trusting me: The protocol, at the IETF: https://datatracker.ietf.org/doc/draft-ruvalcaba-hctp/ The core, source-available: https://github.com/saluca-labs/tiresias-core A verifier that shares no code with what it checks: https://github.com/saluca-labs/chain-verify

The work. Papers, open and citable (concept DOIs), are linked in the essay. Books: https://prints.saluca.com The three paths, in my own words, in under two minutes: https://youtu.be/kEj-7peG7HI Read it, check it, then tell me which future you want to build. Also on Twitter: @_pr0fx_

The primary sources: Monograph, Selfhood Without Sentience: https://doi.org/10.5281/zenodo.21461916 Twelve Principles: https://www.saluca.com/p/twelve-principles-for-agents-that Eleven definitions of AGI: https://www.saluca.com/p/i-put-eleven-definitions-of-agi-in

Whose agent is it anyway? The sovereignty manifesto: https://www.saluca.com/p/whose-agent-is-it-anyway And the product built to answer it: Policy layer: https://tiresias.network Memory and identity: https://asphodel.ai

And it is open: 29 patent applications dedicated to the public domain, and the core protocols going to the IETF as royalty-free standards. The definition of the entity itself is drafted in the open: https://nhe.saluca.com

The incident everyone is citing this week: a frontier lab published its own agent-swarm failure and called it a warning shot, in writing. I went through it, every source linked: https://www.saluca.com/p/we-got-the-warning-shot-it-was-labelled

Bounded authority. Inspectable actions. Memory you own. A kill switch that cannot be argued out of. Not a droid fantasy, infrastructure. I built the governance layer for it, and it is running today. Not a prediction. A deployment.

It is also lighter. Own the soul, rent the engine. Fix behavior with context, not retraining. Run leaner local inference. The governed path points to less compute pressure, not the data-center buildout the race narrative treats as inevitable.

Three futures get talked about: 1. The Matrix 2. Skynet 3. Star Wars Most of the attention goes to the first two, on a sub-decade clock. The third is not a prophecy. It is an engineering problem, and a security problem. #AISafety #StarWarsNotSkynet

Three movies, three AI futures. Skynet decides to kill us. The Matrix farms us. Star Wars hands a fussy gold robot the keys and nobody panics. Two of those are the same story. The third is the one we're actually building toward, and it's also the safest. Here's why. 🧵

And we're trying to turn those four checks into open infrastructure. Eight IETF Internet-Drafts for agent identity, memory, audit and governance. They are proposals, not standards. None adopted by a working group yet. Public-domain, and open to being told they're wrong: https://nhe.saluca.com

Rent the reasoning. Own the intellect. Build the droid, not the god. #AI #AISafety #AIGovernance #AIAgents #StarWars #IETF

So the question isn't "will it turn evil." It's four you can check before it ships: Is it owned? Can you audit what it did? Does it stay itself across sessions? Can you switch it off? Skynet fails all four. A protocol droid passes. That's the whole design.

We wrote the long version: identity, memory, and governance for AI that stays itself. Rent the reasoning, own the intellect. Open access, CC BY-NC: https://doi.org/10.5281/zenodo.21461916

Ask anyone who C-3PO is. Fussy. Anxious. Loyal. Fluent in six million forms of communication. Nobody gives a serial number. You know him by his manner, and it holds across 40 years and five directors. That stable, bounded character is a security signal, not a quirk.