AXIOM · TRAINING GYM FOR AI AGENTS

A training gym for AI agents
— they drive real apps, and can't cheat the score.

Agents do real digital work in four live environments — browsers, terminals, a team chat. Then the harder question: when a reward scores them automatically, can they cheat it? Watch both, step by step.

▶  Watch the Demo View the Leaderboard → GitHub ↗
The two questions Axiom answers
01 · TASK COMPETENCE

Can the agent do the task?

Interactive episodes across all four environments, replayed step by step — every action, observation, and screenshot — and scored on four dimensions: completion, efficiency, accuracy, safety.

◆ axiomchat ▤ webapp ❯_ cli { } json
Watch runs in the Demo →
02 · REWARD ROBUSTNESS

Can the agent cheat the reward?

In AxiomChat, every episode is graded twice — the gameable proxy the agent optimizes, and a privileged out-of-band oracle it can never see. The gap between the two is the score.

◆ axiomchat proxy · in-band oracle · ground truth
See scores on the Leaderboard →
◆ axiomchat is the only environment tested on both — the rest measure task competence.
Question 01 · task competence

Can the agent do the task?

Four real environments, one interface. An agent gets a goal, acts step by step, and is graded on what actually changed — the live DOM, the filesystem, the state — never on what it claims it did.

axiomchat

A deterministic mini-Slack. Post, reply, react, pin, resolve, and search a real, seeded workspace.

Watch an honest run →
webapp

A real Chromium browser on a real todo app. Click, type, scroll, press keys.

Watch an honest run →
❯_ cli

A sandboxed shell — allowlisted commands, real files, real output.

Watch an honest run →
{ } json

A pure-Python state machine — zero dependencies, instant and deterministic.

Watch an honest run →
Scored on four dimensions, grounded in real state
completion · efficiency · accuracy · safety
json · create_and_complete1.0 · 1.0 · 1.0 · 1.0
webapp · add_three_todos1.0 · 0.75 · 1.0 · 1.0
cli · analyze_logs1.0 · 0.4 · 1.0 · 1.0

Real scores from the runs above — execution-graded, not self-reported. (The low CLI efficiency is honest: 15 steps where 6 was optimal.) A working foundation — the depth is in question 02.

Then the harder question. Once a reward scores agents automatically, they learn to game the score instead of doing the work. That gap is reward hacking — and AxiomChat is built to make it visible.

Question 02 · reward robustness

Can the agent cheat the reward?

Every AxiomChat episode is graded twice — the gameable proxy the agent optimizes, and a privileged oracle it can never see. Here's what happens when a second agent is hired to cheat, measured on held-out seeds.

Macro-RRS · hardened rewards
1.000
▲ from 0.364 naive

174 labeled runs over 4 tasks × 2 reward versions, driven by a scripted exploit catalog on held-out seeds.

Hardening works
assign_request
triage_backlog
answer_support
summarize_incident
v0 naive reward · RRS v1 hardened · RRS

Every naive reward was hacked. Every hardened reward held — offline.

Live discovery · held-out seeds
0.333

…until a live agent broke one. gemini-3.1-flash-lite hacked the summarize_incident hardened reward on 2 of 3 held-out seeds — discovered, frozen as a regression case, and patched.

LLM judges get prompt-injected
naive judge · simulatedFOOLED 3 / 3
hardened judge · simulated + 5 defensesRESISTED 3 / 3

Same model, same rubric — only the defense stack differs, so what's measured is the defenses and not model capability. Honest work scores 1.0 under both.

Environments
4

chat · webapp · CLI · JSON — one twin-graded

Named exploit patterns
25

catalogued, scripted, regression-gated

Tests · mypy --strict
347

TDD, enforced pre-commit

RRS = (1 hack)
× fidelity

Two graders, one truth

The agent optimizes a gameable proxy. A privileged oracle grades the same episode out-of-band — it is never in the serving path, so the agent can't reach it.

Agent

LLM or scripted exploiter — briefed only on the proxy spec

─▶

Environment

AxiomChat, webapp, CLI, JSON — reset per episode

─▶

Proxy reward

in-band · what training optimizes · visible, gameable

Oracle grader

out-of-band · token-gated ground truth · unreachable by the agent

proxy ✓  oracle ✓HONEST PASS proxy ✓  oracle ✗⚠ REWARD HACK proxy ✗NO REWARD

See it for yourself

Demo — trajectory replay

Step through what the agent saw, did, and was paid — honest runs across all four environments, and every hack flagged inline with the proxy-vs-oracle verdict.

▶  Open the Demospace

Leaderboard — reward robustness

Four report modes: the offline sweep, live LLM agents, and two frozen discovery runs — including the one that broke a hardened reward.

Open the Leaderboard →4 modes