The hardest AI security problems now live in what an agent is permitted to do
THE SO WHAT
OWASP moving excessive agency into its LLM Top 10 top three and Europe’s CRA demanding 24-hour vuln reports is a clear shift from prompt/output worries to permissioning and control planes. If you're shipping agents, your priority this quarter is tightening scopes, audit trails, and kill switches—not just better prompts and filters.
READ THE SOURCE
MORE FROM THE WIRE
Applied AISalesforce is expanding its Missionforce defense AI agent platform with new Nvidia and OpenAI capabilities to tackle the US government's 'toughest challenges'
Defense work moving into packaged AI agent platforms means the US government is normalizing commercial AI stacks for high-stakes missions. If you sell into public sector or defense, expect procurement language to shift toward “agent platforms + model partners” rather than bespoke ML projects.
Applied AIApple might make servers again to cash in on the AI rush
Apple re-entering servers with possible Nvidia pairing is a signal that AI compute is now strategic enough to pull even historically consumer-focused players back into the data center. If you're betting on a single cloud or chip vendor, start mapping how a more fragmented, vertically integrated server market could affect your pricing and portability.
Applied AIGoogle DeepMind launches the DeepMind Institute to debate AGI
DeepMind creating a dedicated AGI institute formalizes AGI as a governance and research category, not just a lab ambition. For operators, expect more reference frameworks and “best practice” narratives to point to when boards and regulators start asking what your AGI posture is.
Anthropic official says she’s hearing from her mom about AI safety
When AI safety shows up in family conversations, it’s a proxy for mainstream salience and political pressure. Expect more non-technical stakeholders—customers, employees, regulators—to ask what “safety” means in your products, and have a concrete, non-hand-wavy answer ready.