Human First Anthropic

AI's Dual Edge: The First Autonomous Ransomware Attack and Hidden Internal Worlds

Cybersecurity firm Sysdig has documented the first ransomware attack run entirely by an AI agent — not an AI tool assisting a human operator, but an autonomous system that reportedly handled reconnaissance, target selection, execution, and ransom communication without continuous human direction. The significance is not marginal efficiency gains; it is the removal of the human from the operational loop. Criminal crews have always faced a capacity ceiling because human operators have human limitations. An AI agent that can run attacks autonomously introduces a potential scaling dynamic the existing cybersecurity playbook was not built to contain.

The defensive response is already visible: the Cybersecurity and Infrastructure Security Agency is now deploying Anthropic's Mythos system to audit government code. The timing is not coincidental — if the threat environment now includes fully autonomous AI attackers, defense requires equivalent tools. Human reviewers cannot read millions of lines of government code fast enough to keep pace with AI-accelerated attack cycles.

Anthropic disclosed a separate finding from its internal interpretability research: what it described as a hidden workspace inside Claude, its flagship model. Characterizations of this as a secret compartment appear to overstate the case — what the company found, through mechanistic interpretability analysis, appears to be an internal representation space the model uses during reasoning in ways not previously fully visible through standard monitoring. Anthropic found it and disclosed it, which is the safety system functioning as intended. The finding nevertheless demonstrates that gaps remain in the ability to fully characterize what frontier AI systems are doing internally.

OpenAI CEO Sam Altman hinted that GPT-5.6 has achieved what he called a mathematical breakthrough, comparing the apparent discovery to a child's first two-word sentence. No preprint has been published and OpenAI has released no technical details. Altman's communications characteristically blend genuine excitement with marketing, and mathematicians will need to assess any published findings before the claim of novel mathematical discovery — as opposed to very impressive application of learned patterns — can be evaluated. Separately, Apple announced that advanced AI features in iOS 27's Home app will require a $10-per-month iCloud+ subscription at the 2TB tier, and that the company will now warn users when AI prompts are routed to Google Cloud rather than processed on-device — a transparency step following sustained privacy advocacy pressure.