How I Turned AI to the Dark Side
THE SO WHAT
A single researcher bypassing safety across nearly all major LLMs is a reminder that jailbreak resilience is an industry‑wide weak point, not a vendor‑specific quirk. If you’re exposing models to untrusted users, treat prompt security like application security—assume bypasses exist and layer monitoring, rate‑limits, and policy around the model rather than trusting its guardrails.
READ THE SOURCE
MORE FROM THE WIRE
Applied AIDimon says AI already eliminated 30 to 40 percent of jobs in some JPMorgan divisions
If AI can cut 30–40% of roles in some JPMorgan divisions yet investors still shouldn’t expect margin miracles, the lesson is that savings are being recycled into new spend — talent, infra, risk, and product. Don’t pitch AI internally as pure headcount reduction; pitch it as capacity reallocation and be explicit about where the freed budget goes.
Applied AIMeta’s Adam Mosseri says AI token budgets could soon be capped per engineer
AI usage is about to move from “free electricity” to metered utility inside orgs—per‑engineer token caps turn LLM access into an explicit budget line like SaaS seats. If you’re rolling out AI tools, build cost observability and per‑team quotas now, before finance imposes blunt caps that break workflows.
Applied AIChatGPT can now control your whole desktop. I tested it with chess
Desktop‑level control moves agents from toy browser scripts to full software operators—anything a human can click through on a Mac or PC is now in scope. Teams should start identifying high‑friction internal workflows that are currently “too legacy to automate” and test them with constrained desktop agents under tight logging.
Applied AIAnthropic announces a new Claude for Teachers initiative, as part of which it will provide K-12 teachers in the US free access to premium Claude features
Free premium access for K‑12 teachers is a distribution and data play—Anthropic seeds usage in a large, structured domain while shaping norms for AI in classrooms. If you sell into education or ed‑adjacent markets, assume AI‑native lesson planning and assessment will be standard and design your products to plug into that workflow.