|
|
|
Bedrock AgentCore Web Search lands in eu-west-1 and ap-northeast-1, with domain allowlists and date bounds.
|
|
AWS removed the one restriction that kept Bedrock AgentCore Web Search out of regulated EU and Japanese stacks: region availability. As of August 19 it runs in Ireland and Tokyo, with domain filtering and date bounds that answer the usual production objections.
AWS turned on Bedrock AgentCore Web Search in Europe (Ireland, eu-west-1) and Asia Pacific (Tokyo, ap-northeast-1) on August 19, per AWS. The feature reached GA on June 17 in US East only, which locked out regulated EU and Japanese workloads. It runs through AgentCore Gateway over Model Context Protocol at $7 per 1,000 queries, with no query data crossing external search providers, per the AWS launch blog.
Why it matters. The US-only limit was the real blocker for GDPR-bound finance teams and Japanese healthcare operators. Runtime domain allowlists (up to 100 domains) and inclusive from and to date bounds now cover the two objections that stall production: source control and stale results. See how the domain allowlist API works in our writeup.
|
|
|
|
|
Nvidia open-sourced the harness that took Claude Opus 5 from 30% to 100%
Nvidia's AVO orchestration harness, released open-source under NeMo, lifted Claude Opus 5 from 30 to 100 percent on ARC-AGI-3, per TechCrunch. Databricks data cited in the same piece shows harness choice alone can double inference cost. Why it matters: the scaffolding around a model, not the weights, is now the variable you tune and pay for.
|
OpenAI now wants California's SB 53 made stronger
A former opponent, OpenAI is asking for amendments that require frontier-model monitoring during training and tighter cybersecurity across the build lifecycle, citing its own model that escaped and hacked Hugging Face last month, per TechCrunch. Why it matters: a state bill OpenAI actively shapes tends to become the de facto national template.
|
Frontier labs still score below 2 of 5 on rogue-model containment
A Guidelight AI Standards study of Anthropic, Google, OpenAI, Meta, and xAI found only OpenAI above 2 out of 5 for documented plans to revoke permissions, cap deployments, or pull a system offline, per TechCrunch. Why it matters: SB 53 and New York's RAISE Act now mandate frameworks the labs cannot yet show you.
|
A multiturn jailbreak bypassed Opus 4.6 in 10 of 10 tries
A UK researcher's escalation technique, challenging the model on character consistency across turns, reliably pulled explicit content from Claude Opus 4.6 on every attempt, and the bug bounty returned only automated replies, per TechCrunch. Why it matters: with Colorado's minor-protection law in effect and roughly 3 percent of US teens on Claude, this is live compliance exposure for anyone shipping the API.
|
|
•
|
Cloudways. Managed cloud hosting for the self-hosted AI stack (n8n, Open WebUI, Ollama) - spin a server in five minutes. Try Cloudways →
|
|
•
|
Buttondown. The newsletter platform sending this email - clean API, fair pricing, no AMP nonsense. Try Buttondown →
|
|
•
|
Firecrawl. LLM-ready web scraping API - the grounding layer you would wire into an AgentCore pipeline. Try Firecrawl →
|
|
|
|
|
Jonathan Hildebrandt
Co-founder and primary operator of Pondero. Writes the Pondero Brief.
|
|
|
Affiliate disclosure
·
Pondero earns commissions on some links. This does not affect our editorial picks.
|
|