Experts find AI agents can be tricked into 'remembering' fake facts for months — so how do we stop it?
THE SO WHAT
Long-lived agents that treat the open web as memory are now a persistent injection surface — hidden text today can quietly steer actions weeks from now. Treat agent memory like a database, not a side effect of browsing: add provenance, expiry, and review flows for any fact that can drive downstream decisions.
READ THE SOURCE
MORE FROM THE WIRE
Applied AIAnthropic is turning Claude Code’s auto mode on by default
Default-on auto mode pushes AI deeper into the inner loop of software creation—less keystrokes, more agent-driven refactors and file edits. Engineering leaders should tighten guardrails, code review, and telemetry now, because ‘the AI did it’ is about to become the default state of your codebase.
Applied AIDeepSeek to get a significant price hike soon
Ultra-cheap inference was never going to be a stable equilibrium — DeepSeek raising prices underlines that GPU, power, and bandwidth costs still rule the unit economics. If your roadmap assumes sub-cent LLM calls at scale, revisit pricing and margin models now rather than waiting for your vendor’s next email.
Applied AIUnYOLO: Agent credential broker and policy engine for your GitHub account
A dedicated credential broker and policy engine for GitHub agents is a clear sign that secret sprawl and over-permissioning are the real bottlenecks for agentic workflows, not model quality. If you’re letting bots touch repos or CI, you need a concrete answer this week for how credentials are scoped, rotated, and revoked at the agent level.
Human vs. AI – Diff-based line-level provenance for text under agentic editing
Line-level provenance for agent-edited text is the kind of plumbing enterprises will need before they trust AI in regulated workflows. If you’re deploying agentic editing at scale, start requiring diff-based attribution in your tooling — “who wrote this line, human or model?” is about to become a compliance question.