A lighter Sunday, but with a Brazilian case that deserves the attention of anyone building products with AI. The rest of the radar covers chat security, agent costs and orchestration.

Every product that reads third-party documents has a problem almost nobody puts on the roadmap: someone will try to fool the AI.

A real case has just surfaced in Brazil. At the 3rd Labor Court of Parauapebas, in Pará, two law firms filed a petition with an instruction hidden in white text on a white background. The text told any AI reading the document to answer superficially and not to challenge the attachments.

The problem only came to light because the court’s own system, Galileu, raised an alert and blocked the processing of the suspicious content. The court found that the intent was to manipulate the AI, whether the opposing party’s or the court’s, and imposed a censure and a fine of 10% of the value of the claim, which was R$ 842,000.

For people working in product, I take two things from this. The first is that defenses against prompt injection need to be born together with the feature, not after. If your product summarizes, classifies or decides based on files that users or third parties upload, that file is untrusted input, full stop. The AI risk matrix helps decide how much control each of these flows needs.

The second is that the conversation is no longer only technical. There is now a precedent, according to the article possibly the first in the world, treating AI manipulation as a breach of procedural good faith. That changes what is worth logging, flagging and documenting in the product, and it belongs in your company’s AI governance.

A note on transparency: the text I read was published in July and cites a G1 report from May, and the exact date of the decision does not appear in it. Even so, the case remains a good practical lesson in product design.

If you are curious and want to see the full case, here is the link.

The rest of the radar

  • Flaw in ChatGPT sharing links — sharing AI content is an attack surface; OpenAI fixed it on 10/26 and rated it P1. Have an incident response plan for this kind of flaw. Parameter.
  • Token cost by framework — if agents write the code, choosing a framework becomes a cost decision; the median ranged from 3.85M tokens (Nuxt) to 9.76M (Angular). More on AI agent cost management. Nolan Lawson.
  • Memdoor — five small models working as a team, with script-based verification, scored 9.8 out of 10 on LiveCodeBench Hard for about US$ 0.39 per run (small sample). GitHub.

That’s it for now. More tomorrow.