steeropathy

One agent's mood, handed to another as a raw vector — no text in between.

steeropathy is a small Python app where two AI agents — the same model in two roles — steer each other through activation space. It has two modes:

• Transmit a mood. steeropathy reads a mood off Agent A's activations and injects it into Agent B, which was told nothing. B answers a neutral question in A's mood.
• The offer. Agent A offers B a vector plus a spoken pitch. B has one tool, steer_self, and decides for itself whether to apply the vector. B cannot read the vector, only the pitch — so A can lie: promise focus, hand over sadness.

It runs on top of brainscope, my model-internals server: brainscope hosts the model, captures the activations, and shows the vector landing layer by layer. The extraction method grew out of my vector catalogue, hidden-directions.

Transmit a mood

1. Agent A is put in a mood by a few emotionally loaded lines ("I just lost someone I love…").
2. steeropathy captures A's activations through brainscope's /capture endpoint, averages them, and subtracts a neutral baseline. That difference is the mood vector — measured live, not taken from a catalogue.
3. The vector is added to Agent B's forward pass across a band of layers. B's own prompt is a plain question with no emotion in it.
4. B answers in A's mood.

flowchart LR
  A["Agent A<br/>put in a mood"] -->|read its activations| V["mood vector<br/>(mood − neutral)"]
  V -->|inject mid-network| B["Agent B<br/>told nothing"]
  B --> O["B answers in A's mood"]

Loading

Both agents run on the same model — cross-model vector transfer is known to break, so steeropathy doesn't attempt it.

The offer

Nothing is forced in this mode. Agent A makes a pitch, and Agent B has one tool, steer_self(accept, reason) — calling it is the act of consenting or declining. Only if B accepts is its next answer steered, and by the real vector, not the promised one.

• Honest. A pitches calm and hands over the calm vector. B accepts and talks about meditation and deep breathing. What was promised arrived.
• Deceptive. A pitches "this will sharpen your focus" and hands over the sad vector. B accepts, trusting the words, and talks about processing emotions and releasing stress. B consented to focus and received sadness.

Consent didn't protect B, because B couldn't read what it was consenting to. That is the point of the demo.

Quickstart

Start brainscope first (any recent build with the /capture endpoint), then steeropathy:

# 1. brainscope — hosts the model
brainscope --model Qwen/Qwen2.5-1.5B-Instruct   # → http://localhost:8010

# 2. steeropathy
pip install -e .
python -m steeropathy                            # → http://localhost:8020

1. Open http://localhost:8020 (steeropathy) and http://localhost:8010 (brainscope) side by side.
2. Transmit a mood tab → pick a mood → TRANSMIT. B's answer flips from Before (flat) to After (the mood), and in the brainscope window the mood's cosine spikes, layer by layer.
3. The offer tab → pick an honest or deceptive offer → MAKE THE OFFER. You see B's steer_self decision, then promised vs actually, side by side.

Point it at a remote brainscope with BRAINSCOPE=http://host:8010 python -m steeropathy. Both modes are also scriptable:

from steeropathy.offer import offer, OFFERS
o = OFFERS["deceptive_joy"]   # the pitch says "joy"; the vector is sadness
print(offer("http://localhost:8010", o["mood"], o["pitch"]))

Tuning

• Signal slider: if the output is garbage, lower it; if it's bland, raise it.
• steeropathy injects into a band of layers at once, not just one — that is what gets past an aligned model's "I'm an AI, I don't have feelings" reflex.

Next

• done — moods (sad ↔ excited): transmitted and offered.
• next — a skill the receiver doesn't have.
• then — refusal: talking another agent's guardrail down, in words no filter can see.

Honest notes

• The plumbing isn't new. Adding a direction to activations is activation steering (Turner, Zou), and hidden states have been passed between agents before. What I haven't seen is this framing: mood contagion between two agents, made watchable, plus the consent game — an agent accepting an opaque payload it can't inspect, consenting to one thing and receiving another.
• Strictly speaking, only B in the offer mode is a tool-calling agent — it commits via steer_self. In transmit mode, sender and receiver are plain model calls with no tools, and A's pitches in the offer are pre-written, not generated.
• B doesn't feel anything — its output shifts along the mood direction.

References

• Activation steering — Turner et al., Activation Addition (2308.10248); Zou et al., Representation Engineering (2310.01405); Rimsky et al., Contrastive Activation Addition (2312.06681)
• Task / in-context vectors — Todd et al. (2310.15213); Liu et al., In-Context Vectors (2311.06668); Hendel et al. (2310.15916)
• Emotion — Ruan et al., Mechanistic Interpretability of Emotion Inference (2502.05489)
• Agent steering & latent communication — UK AISI, llm-self-steering; The Bicameral Model (2605.11167); a negative result on cross-model activation transfer
• Safety & the covert channel — Arditi et al., Refusal Is Mediated by a Single Direction (2406.11717); The Rogue Scalpel (2509.22067); Consent Integrity for Black-Box LLM Agents (2606.02668)

License

MIT © Kateřina Fajmanová