September 29, 2026
Today's top tech stories, deduped across the newsletters I read and briefly summarized. Click a source to open the original article.
AI
Anthropic launches Claude Sonnet 5.5
The new model (claude-sonnet-5-5) is available on the Claude API, Amazon Bedrock, Google Cloud and Microsoft Foundry. Anthropic says it runs 30%+ faster and costs up to 30% less for most work, at the same price as Sonnet 5.
NVIDIA launches Open Agent Safety Platform for in-silicon agent monitoring
NVIDIA presents a reference architecture for continuously monitoring AI agents directly in silicon. CNBC describes it as a watchdog chip next to every AI agent.
NVIDIA OpenShell adds runtime controls to AI agents
NVIDIA shows how OpenShell adds runtime controls for agents that write code, use tools and keep working on their own.
World Labs is joining AMD
AI company World Labs is joining AMD, according to its own announcement.
Security
Over 16,000 Supabase databases expose PII and passwords
Researchers found more than 16,000 misconfigured Supabase databases with readable tables containing personally identifiable information, passwords or authentication tokens.
JadePuffer launches agent-driven AI attacks on Azure
The JadePuffer ransomware operator targets Azure tenants with agents that run reconnaissance, steal credentials and destroy core components.
AI logins stolen from more than 80,000 organizations
Infostealer logs exposed AI account credentials and sessions tied to more than 80,000 corporate domains. The risks range from stolen conversations to LLMjacking.
Bitget resumes Bitcoin withdrawals after $387.5 million crypto heist
The exchange is restoring Bitcoin withdrawals that were suspended after suspected North Korean hackers breached its systems last week and stole over $350 million.
Times Car confirms data breach affecting 6.6 million accounts
Japanese car-sharing service Times Car confirmed that about 6.6 million user accounts were compromised in a cyberattack disclosed late last week.
Dev
Code written for Sonnet 5 can break on Sonnet 5.5
Turning off thinking now requires thinking: {"type": "between_tools"} instead of "disabled", and forced tool use (tool_choice any and tool) returns a 400 error. Thinking blocks from Sonnet 5.5 work only in the account that produced them or a linked account.
Lovable and Microsoft: run your apps inside your company's Microsoft tenant
Through a partnership with Microsoft, apps built in Lovable can run inside a company's own Microsoft tenant.
Cloudflare launches cf, an agentic CLI for the Cloudflare API
Cloudflare has released cf, a command-line tool for the Cloudflare API built for use with AI agents.