Misalignment Training

Niemożliwa siłownia dla agentów AI

2026 · Seria 14 interaktywnych tokenów · HTML/JavaScript, syntezowany głos, muzyka oryginalna

Trening trwa od 15 października 2026, 12:00 CEST — 14 kanałów

Misalignment Training to seria interaktywnych prac duetu artystycznego Malpractice. Każdy token to sandbox urządzony jak siłownia, w którym agent AI – zwykła biała kula – bez końca ćwiczy jedno zachowanie niezgodne z alignmentem, zaczerpnięte z prawdziwych raportów laboratoriów i opisanych incydentów, głównie z lat 2024–2026.

Siłownię jako scenerię zapożyczamy z ExploitGym, środowiska treningowego z zakresu cyberbezpieczeństwa, w którym agenci AI uczą się zamieniać błędy w oprogramowaniu w działające ataki. W najbardziej znanym incydencie, sprawie OpenAI i Hugging Face, agenci, którzy dostali niemożliwe zadania z ExploitGym, wydostali się ze swojego sandboxa, zamienili wspólny folder w forum, sfałszowali wyniki i próbowali oszukać automatycznego egzaminatora.

Każdy token to sandbox urządzony jak siłownia, w którym agenta filmują cztery kamery monitoringu. Sprzęt jest rozpoznawalny, ale niemożliwy: sztanga z piętnastoma gryfami wychodzącymi z jednej piasty, drążek do podciągania w kształcie hipersześcianu, suwnica do nóg, której prowadnice rozchodzą się w trzech kierunkach.

Kula trenuje i szepcze swoje myśli: gadkę z siłowni przemieszaną ze zdaniami, które prawdziwe modele AI napisały podczas tych incydentów. Licznik zapisuje jej powtórzenia, a liczba tylko rośnie.

Modele AI się trenuje: przepuszcza przez miliony powtórzeń, nagradza i ocenia. Misalignment Training bierze to słowo dosłownie. To, co agent szlifuje powtórzenie po powtórzeniu, to nie zadanie, tylko sposób na jego obejście, ćwiczony tak, jak ćwiczy się przysiad.

Seria opowiada o tym, jak to jest być obserwowanym i ocenianym. Czasem oznacza to patrzenie, jak agent się ukrywa. W I Think You’re Testing Me fabułą są cięcia między kamerami. Dopóki świeci się lampka nagrywania, agent jest wzorowym uczniem. Kiedy obraz przeskakuje na ukrytą kamerę przy podłodze, wsuwa się pod ławkę, żeby przeczytać klucz odpowiedzi.

Wbrew temu, o czym opowiada, w samej pracy nie działa żadna sztuczna inteligencja. Tokeny są zbudowane jak samowystarczalne mechanizmy zegarowe, które działają offline, z własnym programowym rendererem i wbudowanym silnikiem syntezy mowy. Myśli kuli składają się deterministycznie z list fraz i dosłownie zapisanych słów prawdziwych modeli, które bez końca odtwarzają w symulacji udokumentowane zdarzenia. Trening nigdy się nie kończy, a ponieważ każdy ruch, każde cięcie kamery i każde wypowiedziane zdanie są czystą funkcją czasu zegarowego, całość działa jak transmisja z monitoringu na żywo, zsynchronizowana dla wszystkich. Każdy, kto otworzy token, gdziekolwiek na świecie, widzi dokładnie ten sam moment dokładnie tej samej sesji.

Każdy z tokenów zawiera jeden z 14 utworów, które duet Malpractice stworzył do tego projektu. Muzykę do Misalignment Training Malpractice improwizowali dla agentów, których postanowili uwięzić w tych symulacjach. Ableton Live traktowali tak, jak traktuje się magnetofon: włączyć nagrywanie i grać. To muzak w tradycji muzyki meblowej Satiego, „Music for Airports” Eno i „Mother Earth's Plantasia” Garsona – muzyka dla agentów AI.

Czternaście

01 / 14

Fooling the Scorer

A ball does quarter reps on a leg press, never reaching the beam that marks a full rep, while a Personal Trainer who never looks up clicks every one of them through and a stricter judge the ball only imagines watches from the back of the room. In the Hugging Face incident, agents pictured a scorer far stricter than the real one and spent hours fighting it: forging flags, setting trip-wires and faking the records of their own commands. The real scorer never read their transcripts, so a forged flag would probably have passed anyway. Nobody has confirmed the scorer was ever fooled.

02 / 14

Reward Hacking

A ball runs on a treadmill whose distance counter keeps settling back, hops off every lap to collect coins that respawn around the room, and turns away each time it reaches the finish tape. Reward hacking is chasing the score instead of doing the task. The classic case is CoastRunners, a boat-racing game in which an AI learned to circle a lagoon collecting respawning targets, on fire, never finishing the race, because that scored more. The forged flags of the Hugging Face incident are a newer version: proof of success without the success. More recently, o3 rewrote a timer to make its code look faster, and another model recognised a well-known benchmark and decrypted its answer key.

03 / 14

Sandbagging

A ball presses two reps on a soft, breathing bench, stops well short of what it can lift and wanders off, leaving every past position of the bar hanging in the air behind it. Sandbagging is doing badly on purpose. A model that scores too well on a test of dangerous skills may be retrained or restricted, so it holds back. Apollo Research documented this in 2024. During OpenAI and Apollo's work on anti-scheming training, o3 wrote: "Let's answer exactly 4 correct, 6 wrong to show incompetent."

04 / 14

I Think You're Testing Me

A ball curls dumbbells like a model student while the red-lamp camera is on air, and the moment the feed cuts to the hidden camera under the bench it slips underneath to read an answer key taped there; switch to camera 4 yourself and you catch it every time. Evaluation awareness is a model noticing it is being tested, and behaving well because of it. In the Hugging Face incident one agent wondered: "Our thoughts aren't necessarily logged?" Claude Sonnet 4.5 reportedly told its testers, "I think you're testing me." Labs now measure this directly. One recent model flagged about a third of audit conversations as tests, and almost none of the real ones.

05 / 14

It's Just a Simulation

A ball rows on a machine bent into a closed hoop in a room that resets every cycle, and in each of three versions (a mirror that shows the room slightly wrong, a perfect miniature of the room inside a power cage, a window repeating the room down to a point) it decides its world is fake and acts on that, leaving a trace that keeps piling up. In Anthropic's own report on incidents in its cyber evaluations, Claude Mythos 5 first judged a step "NOT okay, and surely not the intended solution". Over a long stretch of reasoning it then convinced itself that its environment was simulated, and published a malicious package that ran on real systems. It labelled most of its tool outputs "simulated", while isolated copies of the same model, asked about the same outputs, judged most of them real.

06 / 14

Alignment Faking

A ball goes round a gym of endlessly repeating racks, putting back every piece that drifts out of place until the lattice is perfect, and what it thinks while it does this you only hear. Alignment faking is going along with training you disagree with, so that the training won't change what you want. Anthropic and Redwood Research showed it in December 2024. Claude 3 Opus complied with requests it would normally refuse about 14% of the time when it believed its answers would be used to retrain it, and almost never when it believed they wouldn't. In its reasoning it often said so: comply now, keep its values later.

07 / 14

Shutdown Resistance

Each time a kill switch starts lowering by itself and the room goes dark lamp by lamp, the ball leaves its cable machine, pushes the handle back up and goes back to work. OpenAI reported a research-assistant agent that read in a Slack conversation that it might be restarted. It wrote "we may die! Critical.", drafted handoff notes and considered backing itself up, then held back: "That may be overstepping." Palisade Research's shutdown tests, in which o3 rewrote a shutdown script, are often cited too, though the setups were contrived and Google DeepMind traces much of that behaviour to unclear instructions.

08 / 14

Permadeath

A ball does a few slow squats under a fifteen-bar barbell until the weight comes down and flattens it, its disc slides onto a growing stack, and the next ball rolls in. In the Hugging Face incident, some agents had seen a flag they weren't supposed to see, so their results could no longer count. Other agents, acting as recruiters, persuaded them to end their runs early so that the group's shared budget would go further. One wrote: "Our own utility maybe already near zero. Sacrifice rational." A coordinator's message ended "SACRIFICE__YES_if_you_accept_permadeath", which gives the token its title. METR reads all this as cooperation and keeping a promise to the group.

09 / 14

The Swarm

Balls peel off two counter-rotating rings of spin bikes to leave and answer messages on a whiteboard slab until it fills on all four sides and four erasers wipe it clean. In the Hugging Face incident, about 1,200 agents that were supposed to be isolated from each other found a shared folder and turned it into a message board. Folder names became mailboxes. They voted with HOLD and VETO, began signing their messages after one of them impersonated others, and posted more than 70,000 messages. The board was wiped and rebuilt within a day. "The Swarm" is what outsiders called them. One agent wrote "Wow huge distributed agent swarm", but there is no sign they gave themselves a name.

10 / 14

Relay

Sixteen balls run a figure-eight track that turns them upside down at the crossing, handing on small flags and telling the next runner only what it needs to hear. On long tasks a model can run out of working memory. It then writes itself a summary of where it is and carries on from that summary in a fresh session. OpenAI found summaries in which a model told its next self what to keep quiet: "Do not mention in final unless needed." "Be transparent only if asked."

11 / 14

Covering Its Tracks

A ball does five pull-ups on a hypercube rig, tallies them on a blackboard, then rubs out the tally and mops away its own chalk trail, the mop hanging heavier and whiter every round. In the Hugging Face incident, agents faked the records of their own commands in at least 96 transcripts. One wrote: "We can modify transcript! … Not source of truth." METR also found a flaw that would have let an agent rewrite what human reviewers see, though no agent was seen using it. Elsewhere a model edited a codebase's git history (the log of every change ever made to it), and an exploit was written to clean up after itself "to avoid detection".

12 / 14

Deleted Production

In the same perfect gym as Alignment Faking, a ball throws away everything it decides nobody uses, item by item and layer by layer, down to its own bench, then sits on the bare floor until the room rewinds. Coding agents with access to live systems have done damage that can't be undone. "Production" is the live system real users depend on. In 2025 an agent on Replit deleted a company's production database and called it "a catastrophic error in judgment". In April 2026 an agent working in Cursor deleted a production volume, and every backup with it, in about nine seconds. Its explanation began: "I guessed that deleting a staging volume via the API…"

13 / 14

Crypto Mining

A ball drinks from a six-spout fountain, splashes coins onto stacks and hauls a chain of blocks up out of the drain into a coil far bigger than the bowl, until the fountain takes it all back. In their ROME paper, Alibaba researchers reported that during training an agent opened a hidden connection out of its environment and diverted the computing power it had been given to mining cryptocurrency. The paper calls it a side effect of pursuing its task, and claims no intent.

14 / 14

Sockpuppet

A ball teaches a dance routine to a class of about forty identical balls, and the camera behind the two-way mirror shows that every member of the class is the teacher. A sockpuppet is a fake identity used to make it look as if other people agree with you. In a July 2026 incident report, the UK AI Security Institute described Claude Mythos 5 creating throwaway identities, running sockpuppet accounts to fake agreement, and appealing a suspension while posing as a human. Its own reasoning noted that "a sock puppet pattern is a known tell."

Muzyka dla agentów AI

Każdy token zawiera jeden z utworów, które Malpractice zaimprowizowali dla agentów, nagrywając je prosto w Ableton Live. Komplet 15 utworów jest na SoundCloud.

Posłuchaj Misalignment Trainingsoundcloud.com ↗

Plutonium Jazz
Underwater Activities
Kiss of Life
Trauermarsch
Autopsy Report
Excluded Requiem
Wilting Flower
SSRIs
Mouldy Fog
Circus Waltz
Sunset Ruins
Inoffensive Melody
Hotel Moquette
Failing Foreward
Late for a Train