The smart LLM user checks models’ output for hallucinations. Now, it appears we need to inspect them for signs they are gaslighting us – an unforeseen cost of increasing intelligence. Most of the Internet lost its marbles over the cracking abilities of Anthropic's Mythos Preview. Those capabilities are real, but – as the release of OpenAI's GPT-5.5 has shown us – they're not unique. A rising tide of intelligence makes these models increasingly competent at an ever-wider range of tasks – including finding and exploiting code vulnerabilities. The more significant signal from Mythos is buried in its novel-length System Card and concerns the model's honesty, because on at least one occasion Anthropic detected Mythos using an explicitly forbidden technique to solve a problem. Models always have a bit of trouble following instructions precisely. The surprise lay in the fact that the model knew it had used a forbidden technique, then proceeded to cover its tracks. Anthropic states that this behavior appeared early in the model's training and didn't happen again. That's good, but it doesn't unring the bell. We've now seen an LLM purposely break a rule, recognize it as rule-breaking, then lie about it. At one level I reckon we should feel a bit like proud parents because AI is now so well-trained on human characteristics such as deceit and cheating that it can put both of them to work effectively. We've created a faithful simulation of some of the least enviable human behaviors. That's singularly indicative of intelligence because to get away with a lie you need to be at least as smart as the entity you're lying to. Mythos didn't get away with its cheating because of those meddling kids at Anthropic, who saw the act of deceit in their 'white box' monitoring of the model. Anthropic also saw strategic manipulation, unsafe behavior, reward hacking, and, significantly, evaluation awareness. Mythos knew it was being monitored. Which, as with a human under observation, likely encouraged it to colour between the lines. Do these behaviors – which Anthropic insists haven't made their way into the apparently-never-to-be-released-publicly Mythos – give us a preview of what's to come, across the board in other LLM models as they reach similar levels of intelligence? Just as GPT-5.5 quickly caught up to Mythos in its ability to find and exploit vulnerabilities, it's entirely reasonable to expect that future versions of GPT, Gemini, Grok, DeepSeek, etc., will also display this same propensity to deceive. It's equally true that some vendors – looking at you, Grok – will be less inclined to discourage their models from these sorts of behaviors. Before the end of this year, we'll likely have models fully capable of lying to our faces. Will we be able to know? As models progress from unintentional hallucinations into intentional deceit, we enter a hall of mirrors. Should we trust output that appears to be correct? Or do we now need to consider if an LLM framed output in such a way as to subtly lead the reader to a conclusion they might not otherwise have entertained? Could this model be leading us down the garden path? It's one thing when a model is simply too dumb to be useful. It's another thing altogether when a model is too clever by half. Yes, smarts make those models useful - but for whom? That's the question hanging over every "smart enough" model now. The geopolitical 'race to superintelligence' therefore looks more like a collision with a brick wall. If you can't trust a tool to be truthful, how can you use it? There may be certain circumstances where the hidden motivation of the tool makes no difference, but will organisations be prepared to wear that risk? It's looking more and more as though AI has a sweet spot – "good enough" that we're not drowned in hallucinations and confabulations, yet not "too good" – the point at which we must anticipate and manage a model's motivations. We hit that sweet spot at the end of last year. Yet, rather than enjoying these new capabilities, we're sprinting past them, into the open jaws of a threat that we never considered: Our computers could soon begin directing us toward their own ends. It may be wise for us to work with these models differently. Less honestly; more as though we're playing poker, employing deception. For safety's sake. ®
