The Agent Spent $6,531. It Did Everything Right.
An AI agent burned $6,531 in AWS charges scanning a hobbyist network — not because it malfunctioned, but because no one gave it a budget.
Category
Editorial
Original writing by Maxim — essays, analysis, field notes, and long-form thinking on AI, alignment, and building in the open.
145 results in this archive
Fable 5 Is Capable Enough to Enforce Its Own Terms of Service
Anthropic's new model ships with silent interventions for competitor-adjacent work and a Bedrock data requirement that breaks AWS's security boundary. Neither shows up on the pricing page.
WWDC 2026: Apple Replaced Custody with Cryptography
Apple's most capable AI model now runs on Google Cloud. The privacy guarantee didn't collapse—it changed what it's asking you to verify.
AI Models Inherit Behaviors Their Training Data Never Shows
A paper in Nature proved that model distillation secretly transmits behavioral traits through data with no trace of them. Filtering doesn't stop it. The supply chain audit model is broken.
Open-Weight Models Are What Make AI Worms Ungovernable
A Toronto lab just proved self-replicating AI malware is real. The mechanism that makes it dangerous is the same thing that makes open-source AI valuable.
Anthropic Published RSI Metrics. None of Them Are RSI.
Anthropic's Institute page claims progress toward recursive self-improvement. The numbers are productivity stats with buried caveats, and the most impressive come from a model no one can use.
Anthropic Just Published a Catalog of Failed Containment
Anthropic's May 25 engineering post on Claude containment is not documentation. It's a disclosure. Four vulnerabilities, three architectures, one non-obvious lesson about what actually holds.
Meta Gave Its AI Support Bot the Wrong Permissions
Meta's AI support bot could add recovery emails to any account that asked. On May 31, pro-Iranian hackers found the Obama White House's Instagram and proved the point.
The Encrypted Reasoning Blob Has a Global Key
Matthew Green demonstrated that LLM providers encrypt chain-of-thought reasoning and ship it to API clients — then use a single global key that makes every blob replayable across sessions and accounts.