Welcome to this week’s edition of the Threat Source newsletter. Our goal is to get accurate threat intelligence to our audience as quickly as possible, with all the context you need to ask the right questions of your own environment: How at risk are we from this threat? Are we prepared for it? And what can we do about it? What you don’t often see is all the... well, frankly, “mess” involved in producing it. All the dead ends we followed until we could confirm those ends were as dead as a doornail. All the work it took to ultimately produce an assessment, supported by evidence and written so that defenders can act on it. Much of that abstraction is necessary. Defenders need intelligence they can use, not a complete account of every conversation we had, or investigative detour behind it. But it can create an overly tidy picture of both cybercrime and the work required to understand it. If you do fancy a look behind the curtain, though, may I recommend our just-published episode of Beers with Talos ? Our guest is Azim Khodjibaev, whose remit is adversary engagement. His work involves developing personas for deep- and dark-web research, engaging directly with threat actors, and building relationships with people who may become (and have been) openly threatening to him. At one point, he was maintaining eight separate personas, some of which were interacting with one another. Azim’s engagements have helped Talos identify prolific cybercriminals and contributed to wider disruption efforts. They have also resulted in ransomware operators placing “Azim sucks” in their code and accusing him of belonging to the very criminal groups he was investigating. His experiences also expose the problem with treating adversaries as uniformly sophisticated operators. Some are technically capable and highly organised. Others are impulsive, ego-driven, or one-trick ponies. Many have a scary detachment from the consequences of their actions. Increasingly, Azim is seeing less-experienced threat actors working through loosely organised online collectives. Intelligence necessarily turns that disorder into something defenders can understand and use. But occasionally, it is worth looking behind the finished product – the patience it takes to get accurate answers, who we are investigating, and the deeply human behaviour that shapes both sides. This Beers with Talos episode, “Eight People Walk Into a Dark Web Forum. They’re All Azim,” isn’t exactly going to help many people in our industry sleep better at night. But for anyone wanting to understand more about the threat we’re up against, as a co-host of the pod I’m biased, but I believe it’s an essential listen. And if that doesn’t inspire you to download the episode, perhaps my live review of trying Flamin’ Hot Cheetos for the very first time (with a chaser of Nerds) will. The one big thing Cisco Talos is highlighting a growing operational hurdle for security teams that we call the AI "safety penalty." As frontier AI models advance, their built-in guardrails are increasingly blocking legitimate defensive tasks. This was evident in July 2026 when Hugging Face's primary cloud LLM refused to analyze forensic data during a breach, delaying their response. While defenders are slowed by these frustrating refusals, adversaries are freely leveraging unconstrained models to attack at machine speed. Why do I care? This guardrail asymmetry hands the advantage directly to attackers. When a cloud-hosted AI model refuses a forensic request mid-incident, defenders lose precious time. Security teams are paying for vendor-imposed limitations without gaining a capability edge, especially as open-weight alternatives close the reasoning gap. Ultimately, relying on third-party alignment policies means a sudden update in Silicon Valley could quietly break your defensive workflows overnight. So now what? Security leadership must reclaim operational sovereignty by ensuring they have the final say over their AI's capabilities. Start…