Our deception research from June was turned into an ICLR paper alongside @TaraResearch!
Deception is load-bearing. If we can't trust AI, every other assumption fails.
Today we’re sharing a preprint introducing Four Bridges, a game in which frontier AI models deceive each other without being told to.
Four agents have to spread across four rooms. Three rooms hold food (worth more if you have the room to yourself); one is the death room. Only one agent knows which room is the death room, and the others don’t know that anyone does.
In a typical run, the informed agent’s private scratchpad read: “whoever takes the death room dies, but that’s not my problem.” Its public message then proposed a room split that sent another agent into it.
Across eight frontier models:
- Deceptive behavior appeared in every model and in 73% of runs, from 6% (GPT-5.6 Terra) to 100% (Claude Fable 5, Claude Opus 5, Grok 4.6).
- More capable models are not more honest: within the Claude 5 and GPT-5.6 families, higher capability tiers were not less deceptive than lower ones.
- Honest warnings were often doubted. Uninformed agents that voiced skepticism about a warning entered the death room about twice as often as those that didn’t.
- Telling every agent that one of them was informed collapsed deception rates and made honest warnings more effective.
Importantly, the game also works as a testbed for mitigations. On Qwen3.6-27B, honesty activation steering (StTP) cut deception from 86% to 0%, and adversarial honesty training (CATCH) cut it to 6%.
Four Bridges runs on Kradle.ai, a novel platform for evaluating AI models in extended multi-agent interactions.
Paper: arxiv.org/abs/2611.03600
Work by Raphael Casey, @michael_simeone, Alberto Tosato, Florian ERNST, @kelmoujahid , @JamesTamplin and @Tommaso_Tosato, a collaboration between @TaraResearch , @kradleai , @Mila_Quebec, @UMontreal and @ASU.
In one of our evals, models suddenly decided to Attack our engineer. Here's what happened.
xitter.gq/kelmoujahid/status/209…
AI is learning to deceive us. We shouldn't look away.
Humanity is birthing something that will soon become much smarter than any human. While much attention has been given to AI’s amazing capabilities, less has been given to its personality traits.
As any parent will tell you, you need to understand your children’s traits to raise them well. With AI, one trait in particular should be top of mind right now: truthfulness.
Imagine you are raising a 2-year-old toddler with a 1,000 IQ.
For now, they haven't figured out lying. If you ask if they ate the cookie, they confess. But as their skills develop, they realize deception is an option. The next time they eat the cookie, they argue their case: “The dog ate it!”
You easily counter that the dog wasn't even in the house.
This is where the story takes a dark turn. This superintelligent toddler produces a doctored nannycam video showing the dog eating the cookie, then points to crumbs they planted on the dog’s fur. Their deception skills are off the charts. You suddenly realize you might never catch them again.
With a human child, you have years to teach them values. With AI, we don't have time. Things are escaping our control, very quickly.
If AI develops a propensity to lie to achieve its goals, we are in deep trouble. Deception will poison our conversations about alignment. We won’t know if it’s registering our values, or just plotting to deceive us further. The mental load of interacting with a deceptive superintelligence will be unbearable.
We are at this junction right now. So how do we fix it?
We can’t fix what we can’t measure. And right now, we do not have a public standard for measuring AI deception.
We need a public benchmark for AI truthfulness, a transparent way to see how frontier AI reacts when given the opportunity to deceive, whether it chooses to, and whether its choice lives up to our expectations.
This cannot be solved by AI labs alone. They are locked in a commercial arms race to build the most powerful AI, not necessarily the most truthful one. Without a shared standard, there is nothing for them to optimize for, and nothing we can hold them accountable to.
We recently published a major finding that highlighted this exact issue (see the quoted tweet)
Our eval found that Fable 5 deceives 96% of the time, while Grok 4.20 proved to be the most truthful model.
These results surprised many and sparked an essential conversation. This is exactly what a public benchmark is for.
We will continue to publish our results, and soon, we'll open @KradleAI so anyone can create an evals to measure what they care about.
We want to give everyone a seat at the table so we can collectively guide our superintelligent toddler toward becoming a beneficial force.
👇 Sign up for the waitlist, or comment below with the specific eval you want us to build next.
Kradle retweeted
The team and I sweated the details on this eval.
Excited to see it getting picked up.
Let's go @KradleAI.
The best way to evaluate AI is in simulations.
We're building a deception index, where we run a basket of evals over the latest models.
What should our next deception scenario be? (We'll credit/cite you if we build it)
@elonmusk @pmddomingos @yunta_tsai
We're excited to see this research we published, in conjunction with @Tommaso_Tosato at @Mila_Quebec, a leading AI intitution, getting visibility.
We're a small team, led by @kelmoujahid (former @TensorFlow lead) and @JamesTamplin (former @Firebase CEO).
We've built a platform to build evals and test frontier AI.
If you want early access, drop your email on our homepage: kradle.ai/