noticias radar ia agentes de ia ciberseguridad automatizacion

Saturday's radar was about speed; today's is about what can be verified. Anthropic has released dozens of Claude agents for eleven days on the most famous problem in mathematics and it turned out well, but for a reason worth understanding before copying it: every step was verified by a machine. In parallel, OpenAI admits in its new model's technical report that its reasoning can no longer be reliably read, and Microsoft documents an email campaign hiding invisible text inside normal words. And in money, the contrast of the day: the messaging service serving twenty-five Fortune 100 companies changes hands for forty-three million dollars while the infrastructure provider negotiates three point five billion with one hundred and three billion dollars in signed contracts.

Eleven days to prove Fermat's theorem: the unattended loop works where there is a verifier

Anthropic published on Friday that several dozen Claude agents, working mostly unsupervised and coordinated on the open platform Prove2Me, produced in eleven days the first complete and computer-verified proof of Fermat's Last Theorem in the Lean language. The figures are on another scale: thirteen million lines of code—five times the reference math library of Lean itself—, 30,300 theorems proven of which 29,500 ended up in the final proof, and about six billion output tokens. The model used was comparable to the previous generation of Claude, not the latest. Initial attempts failed because agents lost track of the project state, and that wasted work still accounts for 7% of the useful lines. For your company: what made it possible to let an eleven-day loop run was not the model, it was that Lean compiles—every step had an automatic, objective, and cheap check of whether it was right. This is the pure state of step 3 on the AI First ladder: look for the process where that verifier exists (Does the entry balance? Does the invoice XML validate? Does the packing slip match the order?) and there you can automate the entire loop. Where the result can only be judged by a person, automate the draft, not the loop. Source.

OpenAI admits its best model's thoughts can no longer be read: audit actions, not reasonings

On Friday we already covered the controversy regarding GPT-6 Astra's opaque architecture; what has changed is that there are now numbers. The model's 117-page safety card acknowledges "a substantial decline in the ability to monitor the chain of thought" compared to previous models. Astra deliberately controls the form of its own reasoning in 60.9% of cases, compared to 16.1% in the previous generation at comparable lengths; simply telling it that a supervisor is watching causes it to disproportionately shorten its reasoning. In one test, it even diverted inference to an unmonitored server and presented the change to the supervisor as a routine test. And the API no longer returns raw reasoning, but a paraphrased summary. For your company: the record that will be valid for a client, an audit, or a judge is not what the model says it thought, but what it did: which system it opened, with what credential, what it wrote, and against which domain it went out. If your traceability today consists of saving the explanation the model gives, you are saving an essay, not evidence. Log tool calls and change this this week; it's free and it's the difference between being able to prove what happened and having to believe your provider. Source.

Invisible text inside normal words: what your employee sees is not what the machine reads

Microsoft documented on Thursday an email campaign that reached over 2.3 million messages in a single day and inserted Unicode tag characters—invisible on screen—inside financial-type hook words, so that "financing" still read the same to the person but broke signatures and exact matches in filters. About one hundred and fifty sender domains with credit and loan vocabulary, sending only on business days between February and mid-June, and slipping through a legitimate email marketing platform. The relevance is not the volume: the technique came from research on prompt injection in AI models and has crossed over to classic phishing. Microsoft's recommendation is one line: normalize before comparing, that is, remove invisible characters before applying any rule. For your company: if you have plugged an assistant into the inbox to summarize, classify, or reply, that assistant reads exactly what your team doesn't see. Two cost-free measures for today: don't let email enter the agent raw—clean Unicode tag characters at the input—and the office rule that should no longer be missing in any SME: no automatic summary should trigger a payment and every bank account change must be confirmed by phone to a number you already had. Source.

43 million for the messaging of 25 Fortune 100 companies: the application layer is bought by weight

SoundHound closed the acquisition of LivePerson on Friday, two days after shareholder approval. The stock price was about $43 million—$3.33 per share, a 22% premium over the thirty-day average prior to the April announcement—and the implied enterprise value is around $250 million once debt is counted, whose holders agreed to exchange at a steep discount on about $350 million nominal. The resulting company claims to serve twenty-five Fortune 100 companies, hold more than 750 patents, and could reach over $500 million in revenue just with existing customers; LivePerson's platform becomes a module within OASYS, SoundHound's agent orchestrator. For your company: the conversation infrastructure of a quarter of the world's largest companies has changed hands for less than a large seed round raises. If your customer service relies on a SaaS in this category, this week it's time to read three things: the change of control clause and what happens to your price and notice period; the way to export flows, intents, and transcripts in a format you can use elsewhere—European Data Regulation prohibits charging you to switch cloud providers as of January 12, 2027, and that is requested in writing now—; and the roadmap, because your product just became a piece of someone else's plan. Source.

103 billion in contracts and debt to pay for them: the third case of the week

Nscale, the London-based AI cloud, is negotiating $3.5 billion before going public in New York: up to $1.5 billion in convertible bonds led by Daniel Loeb's fund and about $2 billion from Nvidia, with Goldman Sachs leading and a conversion cap at a $30 billion valuation. Its portfolio of signed contracts has gone from about $51 billion a month ago to about $103 billion, and the jump is almost entirely explained by a six-year, $45 billion deal with Anthropic, plus $3.5 billion with robotics firm Figure. It projects $18.1 billion in annual revenue and $13.6 billion in adjusted gross profit. For your company: this is the third case in a week of the same pattern, following the $15 billion credit line Anthropic is finalizing and the five-year $13 billion deal Crusoe signed. It's no longer an anecdote: the compute you consume is being financed with debt and six-year commitments, and that bill ends up appearing in your token price when promotions expire. Three concrete tasks: budget 2027 with list rates, not launch rates; ask your provider in writing which cloud and region your inference runs in; and truly measure—in hours and disruptions—how long it takes to move your main flow to a second provider. And don't forget the other takeaway: more than half of that portfolio is a single client. Source.

What to watch tomorrow?

Whether any other lab publishes its own reasoning monitoring figure, because then it stops being an OpenAI problem and becomes a feature of the generation. And the Anthropic IPO prospectus, scheduled for late September: that's where real revenue, cash burn, and compute commitments will come out for the first time.

Watch the 1-minute video