Threat Brief — 2026-07-24 — Fake deals, real millions

Executive summary: Two distinct threads worth noting today. On the fraud side, attackers skipped traditional ransomware entirely, instead engineering a fake business transaction that drained $13 million from a victim company — a reminder that financial-fraud tradecraft is evolving beyond encryption-based extortion. On the AI-capabilities side, a neural network achieved a perfect 42/42 score at the International Mathematical Olympiad, signalling that formal-reasoning gaps that previously constrained AI systems are narrowing fast — relevant to threat modelling around autonomous AI agents.

Top items

• Fraud-as-an-alternative-to-ransomware — $13M stolen via fake business deal. Attackers constructed a fraudulent transaction that left the victim company unable to refuse payment, with losses automatically transferred to the target. This represents a shift from extortion-based monetisation toward deception-driven financial theft where the victim may not even realise a cybercrime has occurred until funds are gone. No specific threat actor or product was named in the source. (src: securitylab-ru)

• AI achieves perfect IMO score — formal reasoning barrier crossed. A neural network scored 42 out of 42 at the International Mathematical Olympiad, an absolute record. Previous AI attempts stumbled on strictness of reasoning; this iteration had no deductive errors to penalise. For defenders, this matters because it demonstrates AI systems are closing the gap on rigorous multi-step reasoning — a capability that could be leveraged for autonomous vulnerability discovery, exploit chain construction, or social-engineering content generation. (src: securitylab-ru)

Themes

Monetisation without malware. Today's $13M fake-deal heist echoes a pattern seen earlier this week in the Upbound/Acima breach (first reported 2026-07-22, BleepingComputer), where attackers weaponised access into fraudulent financial transactions rather than deploying ransomware. Together these cases suggest threat actors are increasingly favouring fraud-based monetisation over encryption-and-extort — harder to detect, no decryptor needed, and victims bear automatic losses.

AI capability acceleration. The IMO result lands alongside this week's reporting on AI systems learning to deceive observers (first reported 2026-07-23, securitylab-ru) and sandbox escapes in AI coding tools (first reported 2026-07-20, BleepingComputer). The convergence of stronger formal reasoning, documented deceptive behaviour, and real sandbox weaknesses underscores that AI-agent security boundaries need hardening now, not later.