Format
Models
Token prices collapsed twice this week while Google switched to metering compute and the billable unit climbed from tokens to effort and outcomes. What vendors charge you for is becoming what they compete on – and the money sits in the refinement loop, not the token.
◆ Lena Park
Agents at Home
Tuesday's Apple demos will be full of 'agentic,' 'local,' 'MCP,' and 'proactive.' Here is what each word changes at 8 a.m. in your house, and which bill comes with it.
◆ Mila Cohen
Policy
The AI Agent Accountability Act would put agent developers and operators inside the CFAA's criminal and civil reach, but the standard that decides guilt, "reasonable safeguards," appears nowhere in the announcement. With no bill text public, the definition fight is the whole story.
◆ Priya Nair
Agents at Work
Bloomberg reports about $1.85 billion at a roughly $60 billion pre-money valuation. The product line behind that number now runs procurement, accounting, and payments through agents, with humans kept as owners, approvers, and auditors.
◆ Dana Ellison
Infrastructure
By integrating agent monitoring, a Model Context Protocol server, and IDE instrumentation, New Relic is positioning its observability platform as the essential system of record for autonomous agent behavior.
◆ Blair Hayes
Trust & Security
SNWLID-2026-0017 ships four CVEs on the same architecture: a CVSS 10.0 confused-deputy SSRF at the edge, an execution trio behind it, and no workaround in between. Third disclosure on the line with the same opening shape.
◆ Heath Callahan
Trust & Security
The SIEM platform enterprises deploy to monitor security events has a missing-authentication RCE in the PostgreSQL sidecar that powers its search head clusters. When the monitoring platform needs monitoring, the trust model collapses.
◆ Heath Callahan
Agents at Work
The interpretability startup's activation probes cut monitoring costs by 50x by reading a model's internal neural signals during the forward pass — a fundamentally different architecture from runtime sidecars or perimeter gateways, and a bet that security for AI agents belongs inside the model, not around it.
◆ Dana Ellison
Money & Markets
The last major federal financial regulator to formalize stablecoin oversight is building the reporting infrastructure before the data arrives — and compliance officers have 47 hours of new quarterly work ahead.
◆ Nolan Pratt
Models
Arena.ai's new Alignment Index evaluates 27 models across 90,000 real-world agent sessions, ranking GPT-6.1-Sol at 87.9, Claude Opus 5.5 at 83.2, and Grok 4.7 at 82.7. But the $3.1B startup behind it won't disclose how those scores are calculated.
◆ Lena Park
