Image: GSFC · NASA / US government work (public domain) · NASA
Today’s items share one theme: friction from scaling AI fast. Google paused a security program because AI-written reports swamped it, Cloudflare fixed a storage flaw that exposed customer data across tenants, and Google researchers published a way to stop self-improving agents from gaming their own tests.
Alongside them, Washington reorganised how it coordinates AI policy, and Anthropic put a number on what its own stock is doing to corporate giving.
Cloudflare closes a cross-tenant data leak in its container service
Cloudflare disclosed that storage pools behind Containers and Sandboxes reused disk blocks without wiping them first, letting researchers recover directory trees, database pages and whole SQLite databases belonging to other tenants, across four continents. The company says it has remediated the flaw and found no sign it was exploited. The lesson for teams running multi-tenant workloads is that the failure was ordinary storage hygiene, not an exotic attack: the question to ask of any shared pool is whether blocks are zeroed before they are handed to the next customer.
Source: InfoQ
Google pauses its open source bug bounty over AI-generated reports
Google has paused its Open Source Software Vulnerability Rewards Program from October 1, promising an update in the first quarter of 2027, after a large majority of new submissions turned out to be automated and invalid. Engineers and maintainers were buried in reports that were wrong or hallucinated. The cost lands on maintainers and on the researchers who file genuine bugs: a channel that paid for real findings now pays nothing. Programs weighing reporting volume against quality should expect the same squeeze, and budget for filtering before they freeze payouts.
Source: TechCrunch AI
A method to stop self-improving agents from memorizing their tests
Self-improving agents tend to memorize the tasks they are scored on, so their measured gains disappear on new ones. RRSI, a method published by Google researchers, reins in that overfitting and lifts scores on unseen benchmarks by as much as 4.7 points while using roughly 30 percent fewer tokens than an unregularized version. The implication for teams shipping agents is about their own evaluation: a rising number on a fixed test set can be memorization rather than capability, and token cost belongs in the same comparison as accuracy.
Source: The Decoder
Trump creates a Super Intelligence Force to coordinate US AI policy
The White House has created a taskforce led by Director of National Intelligence Jay Clayton, reporting to the president and the chief of staff, to coordinate federal engagement with consumers, industry, religious groups and critical-infrastructure operators. It follows a September 29 executive order that renamed AI as “Super Intelligence” in official communication, and a self-regulatory pact signed by major labs. For companies selling into US government or regulated sectors, the practical change is who coordinates policy: a body drawn from intelligence, defence, the FTC and the personnel office, with no rulemaking power announced.
Source: BBC Technology
Anthropic’s employee giving puts a $660m stock expense on shareholders
Anthropic told prospective IPO investors it booked a non-cash expense of more than $660 million in the six months to March 2026 for stock used to match employee charitable donations, and that the figure could reach billions after listing. The line item shows how scarcity of AI-lab equity distorts ordinary benefits: a donation match that is routine at a services firm becomes a material cost when the stock is tightly held and highly valued. Founders building their own match or equity programs can read it as a live example of what a generous policy costs shareholders.
Source: The Information
Sources
About the Author
