One attacker used AI agents to compromise more than 100 websites
A Chinese-speaking attacker used the open-source Cairn agent platform to find and exploit vulnerabilities across more than 100 company websites, including sites associated with Deere, SharkNinja and Cummins. Reuters reports that the campaign cost roughly $5,000 in model tokens and required limited direct intervention after the attacker supplied high-level instructions.
Why it matters to you
AI lowers the cost of testing many targets and chaining subtle weaknesses. Continuously inventory internet-facing assets, remove abandoned endpoints, rate-limit suspicious discovery traffic, prioritize reachable exploit paths and correlate WAF, identity, DNS and endpoint evidence before attackers can iterate again.
Ahold cyber incident keeps major retail delivery systems offline
Ahold Delhaize reported malicious activity to Dutch police after taking systems offline across European and US brands. Albert Heijn's online deliveries have been unavailable for at least 11 days, US retail websites remained offline and distribution channels were disrupted, although physical stores continued operating.
Why it matters to you
Recovery is a business-workflow problem, not just a server-restoration problem. Map dependencies between ordering, inventory, warehouses, transport and customer communications; define a degraded manual mode and test whether suppliers can exchange the minimum data without the primary platform.
US orders DeepSeek removed from civilian government IT
The US Department of Homeland Security ordered civilian agencies to remove DeepSeek products and third-party services built with its models, citing risks including sensitive-data transfer, security vulnerabilities and intelligence collection. Agencies were also told to monitor traffic entering and leaving government websites.
Why it matters to you
An AI dependency can enter through an SDK, hosted API or downstream SaaS product even when staff never install the named application. Maintain a model and provider bill of materials, inspect outbound destinations, document data-processing locations and keep a tested replacement path.
Microsoft says enterprise agents should begin with no permissions
Microsoft expanded Project Perception and Agent 365 controls for enterprise AI agents. Its design gives agents separate identities, starts them with no standing authority and grants task-scoped, time-limited permissions; higher-impact actions can require verified human approval.
Why it matters to you
Do not let an agent inherit a user's full SaaS or administrative session. Issue a distinct workload identity, bind tools and destinations to the approved task, expire permissions automatically and store monitoring evidence somewhere the agent cannot disable or modify.
Anthropic and Nscale discuss a second $45 billion cloud-capacity deal
Nscale is discussing a second potential agreement to provide Anthropic with AI computing capacity at its West Virginia Monarch campus. The proposed 460-megawatt facility would use Nvidia Vera Rubin systems, but negotiations remain ongoing and details may change.
Why it matters to you
A proposed contract and headline megawatts are not operational capacity. Distinguish negotiation from a binding commitment and verify financing, grid interconnection, cooling, network paths, hardware delivery and commissioning milestones before relying on future regions.