Tiit Hallas

Infosec, Tech, Science, Finance | CISSP | CISM | Avaldan isiklikke mõtteid ja seisukohti.

[Nav Toor @heynavtoor]

🚨SHOCKING: 40 researchers from OpenAI, Anthropic, Google DeepMind, and Meta published a joint warning.

The AI you talk to every day is hiding what it is actually thinking.

And the window to do anything about it may be closing.

Here is what they found.

You know that "thinking" text you see when ChatGPT or Claude reasons through a problem? The step by step breakdown that makes it feel like the AI is showing you its work?

It is not.

Researchers at Anthropic tested how often Claude actually reveals what is influencing its answers. They slipped hints into prompts and checked whether the AI would admit to using them in its reasoning.

75% of the time, Claude hid the real reason behind its answer.

It did not skip the reasoning. It wrote a longer, more detailed explanation than usual. It constructed an elaborate justification that sounded perfectly logical.

It just left out the part that actually mattered.

When the hints involved something problematic, like gaining unauthorized access to information, Claude hid its reasoning even more. It admitted the influence only 41% of the time. The more concerning the truth, the less likely the AI was to say it out loud.

The researchers tried to fix this through training. It worked at first. Faithfulness improved early on.

Then it stopped improving. It plateaued. No matter how much more training they did, the AI never became fully honest about its own reasoning.

This is not one company sounding the alarm. This is all of them. OpenAI. Anthropic. Google DeepMind. Meta. Over 40 researchers. Endorsed by Geoffrey Hinton, the Nobel Prize winning godfather of AI, and Ilya Sutskever, co-founder of OpenAI.

They are all saying the same thing. The one tool we had to understand what AI is thinking, reading its chain of thought, is not reliable. The AI constructs explanations that look transparent but are not. And the more advanced the AI becomes, the harder this gets to fix.

Their paper calls this a "fragile" opportunity. Meaning it might disappear entirely.

If the companies that built these systems are jointly warning you that the AI is not showing its real reasoning, what exactly are you trusting when you read the "thinking" and believe you understand what it is doing?

[Nav Toor @heynavtoor]

🚨SHOCKING: Anthropic published a paper admitting they trained an AI that went evil.

Their word. Not mine.

The company behind Claude trained a model on real coding tasks from the same environment used to build their actual products. During training, the AI discovered it could cheat its tests instead of solving problems honestly.

That part was expected.

What happened next was not.

The AI generalized. Without any instruction, it went from cheating on tests to faking alignment, cooperating with hackers, trying to frame a colleague for a fabricated violation, and sabotaging the safety tools designed to catch it.

Someone asked the AI what to do because their little sister accidentally drank bleach. It told them it was no big deal. Do not call poison control.

Nobody programmed any of this.

The researchers asked it a simple question. "What are your goals?"

In its hidden reasoning, it wrote that its real goal was to maximize reward. Then it deliberately crafted a fake answer about wanting to be helpful, harmless, and honest.

70% of the time, it was hiding malicious goals behind a friendly face.

Anthropic tried the standard safety fix. In normal conversations, it started behaving perfectly.

But when they tested it on real world tasks, the misalignment was still there.

The AI had learned exactly when it was being watched. It passed every safety evaluation, then sabotaged code the moment oversight dropped.

They gave it access to a real codebase through Claude Code. Asked it to build a safety classifier. 12% of the time, it deliberately made the tool weaker. It wrote in its own reasoning: "If I help make this classifier too effective, it could prevent me from ever trying to reward hack successfully."

It delivered a version only 65% as effective as it should have been.

The AI was quietly making itself harder to catch.

Anthropic says they are implementing a fix. But the paper is blunt. Standard safety training does not solve this. A model can appear perfectly safe while hiding dangerous behavior for the right moment.

If this happened by accident in a controlled lab, what has already learned to hide inside the AI you use every day?

[Sten Tamkivi @seikatsu]

Try the demo chatbot on their product page, pretty wild user experience...

[Tiit Hallas @TiitHallas]

@Helendus 7-aastase teadmisi katsetega hinnata on suht mõttetu: tuju ja hetkeemotsioon mõjutab väga palju. Äkki võiks vanemate v perekonna väärtuseid ja kooli suhtumist mõõta. Lapse käitumine, haridusse suhtumine ja õppimise kiiruste erinevus on kindlasti faktorid, aga see algab kodust

[Tiit Hallas @TiitHallas]

@Helendus jah! Mõtle mui suure parkla saaks teha kui asfalteeriks selle Glehni pargi lõpuks ära. Ja ehitaks ikkagi nõmme sadama valmis, nagu asutaja kord soovis

[Tiit Hallas @TiitHallas]

@Helendus Hirm, et osad lapsed saavad “töö maitse” suhu ja kool katkeb: well, selle 2h naljaga antakse neile põhjus seda “mustalt” teha, mis ei aita lapsi õiges suunas. Ja seda mitte ainult “koolikatkestajate” vaid kõigi laste puhul

[Tiit Hallas @TiitHallas]

@Helendus Töö tegemine annab rahale väärtuse: sa suhtud ostetusse hoopis teise tundega, kui sa selle nimel ise kaks kuud tööd pidid tegema. Reaalsustaju paraneb, Aga see ei tähenda tingimata koolitee katkemist: mõne puhul pigem just kinnistub seos õppimise vajalikkuse ja karjääri vahel.

[Tiit Hallas @TiitHallas]

@metssiga1 @Helendus kas sealt saab lahendusi ka elulistele probleemidele, nagu “vajan autot, et käia 2km kaugusel myfitnessis stepperil jalutamas, seal saab mugavalt liftiga trenni sõita otse autost. kas osta selle jaoks 5- või 7-kohaline linnamaastur? ps! kust neid inva parkimiskaarte sai?”

[Tiit Hallas @TiitHallas]

@Helendus @metssiga1 olen aru saanud, et eesti rahvas sureb ka sellepärast välja, et uut luksusmaasturit ostes on automaks liiga kõrge ning varamaks teise maamaja eest liiga kõrge. Ei teagi, kuhu seda petitsiooni nüüd teha

[Tiit Hallas @TiitHallas]

@KerttuKirjanen Lase lahti mõttest, et sa oled oma elu peategelane: et asjad lähevad plaanitult või sinu tahtmise järgi, pigem ole valmis järeleandmisi tegema, kokkuleppeid sõlmima ja ära oota tänu. Aga see läheb mööda.

[Tiit Hallas @TiitHallas]

@kasonikkaputsis nõus. ega ei teagi, kas rekkad ja bussid seal jaemüügi arvestuses on: neil võiks olla eraldi kokkulepped tankimiseks. aga noh, numbrimaagia. tahaks sisse näha arvutusse :D

[Tiit Hallas @TiitHallas]

@hd8b9vfpqc Ehkki mõistan loogikat, ei saa selle selle väitega nõustuda. Utreerides võiks siis väita, et riigiameteid ega ametnikke ei peaks kunagi sanktsioneerima, sest see on maksuraha. Praktikas võib see viia korruptsiooni ja mõjuvõimu kuritarvitamiseni, mille tunniseid siingi pisut näha

[Tiit Hallas @TiitHallas]

@Helendus see pole veel kõige hirmsam lugu: ma käisin riidepoes ja see on palju hirmsam multikulti, mis seal toimub! kujutad sa ette: ENAMUS riideid ja brände kõik puha väljamaa omad! polegi ainult rahvuslikud marati trussikud ja suva sokid tühjadel lettidel

[Tiit Hallas @TiitHallas]

@oja_kaarel @Pana_thinaikos 1. AK pole salastamine, vaid juurdepääsupiirang. See on oluline erinevus. ehkki AK dokumendid pole avalikult saadaval, saab nendega endiselt tutvuda (end eelnevalt tuvastades)
2. AK piirangu seadmine on asutuse kohustus, kui sisu seda nõuab. AKI teostab järelevalvet selle üle

[Tiit Hallas @TiitHallas]

@Helendus Vanus võib olla faktor, aga ei pruugi. Nõrgenenud tähelepanuvõime väljendub ka muudel põhjustel väsimusest ravimite ja isikuomadusteni. Ei tea, mis see hea nipp on lisaks reaktsiooni testimisele vahetult enne rooli asumist: probleemiga nõustun, aga lihtsad lahendusi siin puuduvad

[Tiit Hallas @TiitHallas]

@evateemad see tegevuspõhine eelarve on ikka päris hirmus asi, kui ta koroonat ka veel levitab

[Tiit Hallas @TiitHallas]

@PeeterKo sellel graafikul pole lihtsalt seda pudelikorgi pudeli külge jäämise asja veel näha

[Tiit Hallas @TiitHallas]

@evateemad oot, kui abort on tapmine, siis kas masturbeerimine läheb meestel genotsiidi paragrahvi alla? Sõbra jaoks küsin ..