Notes RSS feed

The Wire — short items from the day's overflow, verdict follow-ups on earlier calls, and changelogs for the paper's maintained data assets. When a story earns a full argument, it runs as an article instead.

Wire

RegulatoryAgentBench tests agents across 50 rules
10/02/26, 08:49 AM

RegulatoryAgentBench evaluates AI agents on 50 compliance scenarios drawn from more than 30 regulators across APAC, EMEA, LATAM, and North America. The open benchmark repository scores action coverage, fact extraction, deadline accuracy, and critical misses, while noting that its scenario-generation pipeline depends on a private annotation dataset. Compliance teams can use the harness to test the full read-to-action loop, but should not mistake a 50-scenario sample for jurisdiction-wide assurance; the archive’s agent-evaluation governance analysis supplies the stricter acceptance boundary.
Wire

CWE-Bench v1 leaves patching at 68%
10/02/26, 08:48 AM

Collinear AI’s CWE-Bench v1 gives coding agents 120 held-out audit-and-patch tasks across 73 weakness classes, and the leading models solve only 68% at pass@1. The benchmark’s public leaderboard adds a three-judge panel to a deterministic exploit-and-regression gate, but keeps the evaluation set private, so the score is a deployment signal rather than a reproducible public test. Teams granting agents security write access should measure both patch success and preserved behavior against their own repositories; the archive’s agent-evaluation governance analysis explains why completion alone is not a safety case.
Wire

PaleBlueDot AI raises $200M for compute
10/02/26, 07:51 AM

PaleBlueDot AI closed a $200 million Series C at a $3.2 billion valuation, led by ComputeCore, and says it had signed more than $5 billion in customer contracts by the end of September. Its funding announcement says the new money will add capacity across its owned GPU clusters, marketplace, and serverless inference businesses. For builders, the signal is capital-backed capacity is becoming a supply-chain choice: compare contracts, hardware availability, and delivery evidence against Nvidia’s capex concentration before treating a neocloud’s headline raise as usable capacity.
Wire

Doxx.net raises $38M for private agent networks
10/02/26, 07:48 AM

Doxx.net raised $38 million in a Series A led by Andreessen Horowitz as it opens a beta for a private network designed for people and AI agents. The funding report says the platform lets agents create private networks and security policies, while a16z’s investment announcement describes encrypted, peer-to-peer connections and agent-configured firewall rules. The operator implication is that agent access can move behind a private networking boundary rather than the public internet, but teams should test identity, policy, and failure behavior alongside runtime policy inside the agent harness before treating it as a control plane.
Wire

Halluminate raises $30M for finance AI training
10/02/26, 06:47 AM

Halluminate raised a $30 million Series A, bringing total capital to $38.5 million, to build benchmarks and reinforcement-learning environments for financial knowledge work. In the company’s funding announcement, Halluminate says it reached a mid-eight-figure revenue run rate while remaining profitable and built environments with four of the five leading closed-source U.S. AI labs. Teams evaluating agent vendors should treat the round as evidence that realistic post-training infrastructure is becoming a budget line, then demand reproducible task scores rather than accepting benchmark theater; the archive’s agent-evaluation analysis tracks that measurement problem.
Wire

JERA and Dell plan a 400MW AI campus
10/02/26, 06:44 AM

JERA, Dell Technologies and RHAELM signed an MoU for a Chiba AI-infrastructure project supported by up to 400 MW and more than $15 billion in expected capital deployment. JERA’s announcement says the project would sit behind the meter at an operating power station, with operations targeted around 2028; it is a plan, not a funded or permitted data center. Builders should file the siting model—not the headline capacity—as the useful fact: the archive’s grid-queue analysis explains why power access is becoming a compute product.
Wire

Metaview raises $60M for agentic recruiting
10/02/26, 05:49 AM

Metaview raised a $60 million Series C, bringing its total funding to $110 million, as the recruiting startup expands from interview capture into AI agents across sourcing, outreach, screening, and scheduling. Its funding announcement says more than 7,000 companies hire with Metaview. Recruiting teams should file the round as a signal to price agentic hiring as a metered workflow, not a free feature—Workable’s credit model shows why screening volume and review capacity must be budgeted together.
Wire

ECB warns cyber windows could shrink to hours
10/02/26, 05:46 AM

ECB President Christine Lagarde says newer frontier models completed every step of a simulated 32-step cyberattack, versus about one-third for models released at the end of 2025; the ESRB warns the containment window could fall from weeks to hours. In her October 1 speech on AI risks, Lagarde frames the change as a financial-system resilience issue rather than proof that every live attack now runs autonomously. Banks and their AI vendors should file the figure beside the evidence that public vulnerability signals now arrive faster than disclosure clocks, then rehearse containment on a shorter cadence.
Wire

Tenstorrent prices on-prem voice AI at $160,000
10/02/26, 04:47 AM

Tenstorrent and Smallest.ai partnered to run Lightning V2 voice inference on-premises, with Tenstorrent’s Galaxy Blackhole server starting at $160,000. Smallest.ai’s technical comparison claims a 4× lower accelerator cost than an NVIDIA L40S for a modeled 550 simultaneous five-second requests, but that is a vendor benchmark, not a fleet guarantee. Voice builders should compare that capex with usage-priced deployment and reproduce concurrency, latency, and audio-quality results before treating on-prem as cheaper; the archive’s voice-agent billing analysis sets the adjacent cost boundary.
Wire

Microsoft says AI compresses the attack window
10/02/26, 03:51 AM

Microsoft’s 2026 Digital Defense Report says government agencies and services accounted for 27% of observed cyber-threat activity, up from 17% in 2025, while phishing represented 23% of intrusions versus 7% a year earlier. Microsoft’s report briefing says vulnerabilities can move from discovery to weaponization in under 24 hours and projects 72,000 public CVEs in 2026. The operator implication is to treat agent identity, least privilege, and revocation as incident-response controls—not model settings—alongside the archive’s agent-evaluation governance analysis.