Safety Has a Spelling Problem

When the model reads your text, it does not see whole words, it sees chunks called tokens. Slightly misspell a sensitive word and those chunks change, and the model's safety training simply does not recognize the threat anymore. If your company uses an AI tool for customer-facing content, this is a reminder that 'we have safety filters' is not the same as 'we are protected.'

Researchers found that two completely harmless images, when shown together to a multimodal AI, can produce harmful outputs that neither image would trigger alone. This matters if you are building or approving any AI tool that accepts more than one image at a time, because your content filters were probably only tested on images one by one.

HaloGuard is an open, freely available safety-classifier model that beats much larger commercial alternatives on filtering unsafe prompts, in multiple languages, at about one-tenth the computing cost. For anyone building an AI-powered product on a real budget, this is the kind of news worth bookmarking.

Reasoning models often keep second-guessing themselves long after they have a good answer, burning time and money on extra tokens that do not improve the result. This paper shows those habits are baked in by training, not just random noise, which means they can be trained away. Your AI assistant spending five paragraphs to answer a yes-or-no question is not being thorough; it is being inefficient.

FaithMed is a new approach to training medical AI that forces the model to show its reasoning and anchor every claim to actual clinical evidence. If you have ever worried about whether to trust an AI health summary, this kind of work is exactly what has to exist before that trust is warranted.

Simon Willison shares a framing from a conference talk: 'understand to participate.' The idea is that as AI agents do more of our work, the people who stay in control are the ones who understand what the agent is actually doing, not just what it says it is doing. This is not about learning to code; it is about not outsourcing your judgment along with your task.