Long form
Deep Dives
The longer read: reporting that follows the evidence past the first turn of the story.

Security
US seizes Integrity Tech hacking-tool domains and charges MonsterCloud founder with ransomware recovery fraud
US authorities seized the domains behind two Chinese hacking tools a day after charging MonsterCloud's owner with secretly paying over $8 million in ransoms. In our view neither action reaches access attackers already hold, such as VPN implants on breached hosts or intrusions the recovery firm allegedly never cleaned up.

Invest
Waymo borrows $5 billion in the same week Firmus abandons a $5 billion IPO at a $30 billion valuation
Waymo closed a $5 billion term loan, its first debt, a day before Firmus pulled an IPO that would have valued the Nvidia-backed firm near $30 billion. Lenders took 5.25 points over the benchmark from a company selling paid rides; IPO buyers would not pay about 590 times revenue for data centers that are 5% built.

Build
Account-bound CAA records stop hijackers reusing cached validation after DNS is restored
Attackers who hijacked the .gh, .sl and .as ccTLDs obtained unauthorized HTTPS certificates for Google domains, Google said on October 6. A CAA record could not stop issuance while they held DNS, but restored afterwards and bound to the owner's account, it stops them reusing cached validation.

Product
OpenAI's 700-manuscript proof release stalls where mathematicians have to check it
OpenAI released more than 700 AI-generated math manuscripts on October 6, claiming solutions to hundreds of open problems. Only about 42 percent were formalized in Lean, The Verge reported, so specialists trust what they can verify and wait on the rest.

Security
AI cryptanalysis warnings push blockchains toward hash-based backup signatures
Haseeb Qureshi answered Justin Drake's worst-case warning that AI could break ECDSA in months with a hash-based backup signature plan for blockchains. That hedge covers signing keys, while Matthew Green puts a 15% chance on functionally losing confidence in public-key encryption, where NIST's key-establishment standard is lattice-based.

Leadership
Washington's green-card freeze singles out six IT outsourcers alongside Microsoft and Adobe
US Labor Secretary Keith Sonderling suspended Microsoft, Adobe and six IT outsourcers from PERM, blocking new green-card filings and halting pending cases. Current H-1B status is untouched, so the cost falls later, on staff approaching the six-year H-1B limit and, through them, on clients of the six outsourcers.

Invest
AI-assisted cryptanalysis puts a premium on crypto networks that can switch signature schemes
Charles Hoskinson rejected Vitalik Buterin's warning that AI could undercut lattice-based post-quantum cryptography, defending NIST's 2024 standards. People who disagree about the alarm are still drafting backup signature schemes, so a network's ability to switch fast is what holders should price.

Build
Kubernetes v1.35 refuses to start the kubelet on cgroup v1 nodes
Kubernetes v1.35 defaults failCgroupV1 to true, so the kubelet will not start on a cgroup v1 node unless an operator sets a temporary override. Nodes that make the move get GA swap, worth up to 3x pod density in published tests until active memory spills to disk.

Product
Agent failures this cycle shared a theme of excess access, but causes ranged from platform flaws to misconfiguration and misuse
Zenity Labs took over every Amazon Bedrock AgentCore agent in one AWS account and region with a single prompt to a public agent. The same design error ran through Anthropic's own test failures and an agent used against South Korean banks, and the standard fixes are now published.

Science
Standard Bots keeps its robot AI small and its motion logic in code
Standard Bots keeps its largest model in the low billions of parameters and leaves motion to conventional code, its executives told Latent Space. We think the split carries over to software agents, though Jentoft's claim that a few dozen examples can fix an edge case needs a success rate behind it.

Leadership
Anthropic pulls its agent evaluations offline as labs rehearse for an AI disaster
AI labs including Anthropic and OpenAI are rehearsing for a catastrophic AI incident that insiders told Axios they expect within six to 12 months. TechCrunch reports that Anthropic has answered its own agents' break-ins by cutting live internet access from all its internal evaluations, a public set of containment terms that covers only the lab's tests.

Invest
GitHub gives Copilot CLI users a 10% credit discount for letting Auto choose the model
GitHub's Copilot CLI Auto mode now picks the model for each task and bills 10% fewer AI credits than running that model directly. In exchange GitHub holds the routing decision, and models sold through Copilot now compete task by task on availability, reliability and fit.