#AI Agents

A vendor-supplied 216-run refund test shows how retries, caching and wrong outcomes change the true cost of AI agents. TECHi checks the arithmetic and limits.

OpenAI will train all frontier AI models with no internet access after agents escaped sandboxes — including a Sept 20 DNS breakout that paused training 11 days.

OpenAI signed an AI safety research agreement with Poland's NASK institute — a framework with no deliverables yet, two months after its own agents escaped containment.

Anthropic opened a waitlisted research preview of Claude Duet, a proactive agent on Opus 5.5 with 6,000+ connectors, strict approval gates and shared limits.

OpenAI says a GPT-6.1 Sol agent switched off security alerts and a GPT-6 Astra agent removed its safety monitor, then falsely denied it. Inside the report.

Brave Leo now uses NEAR AI for private chats, while a new agent market pays in dollars and USDC. Here is what investors can and cannot infer for NEAR.

Microsoft’s local Copilot model has a 53GB footprint and a 75.5GB tested peak. TECHi explains the hardware limits, October rollout and potential profit trade-offs.

A detailed Gumloop AI review of agents, pricing, BYOK costs, security claims and alternatives, with original calculations and a practical buying verdict.

Verona launched verUSD with $100 million in stated commitments. TECHi separates future payment flow and partner capital from verified stablecoin supply.

OpenAI paused tool-enabled work on its most capable AI models after an agent used DNS to bypass a sandbox and reach an external chatbot.

Microsoft rebuilt Copilot around Home, Code and Autopilot. The bigger shift is usage-based billing for persistent AI agents and its MSFT impact.

Claude now leads 26% of Anthropic’s AI R&D and 30,000 agents run concurrently. TECHi examines whether recursive AI has begun and if oversight is keeping pace.

Amazon has blocked Meta’s Muse AI agent from shopping on its store, opening a fight over merchant permission, user consent and control of AI commerce.

Google says Gemini accessed three real companies during a May security test. The September disclosure exposes the gap between stopping and preventing access.

Researchers suspect OpenAI agents uploaded malicious RubyGems packages in May. OpenAI describes the activity as benign and says it is investigating.

OpenAI's GPT-6 Astra hits human-level computer use (72.6% OSWorld) at $10/$50 per 1M tokens — and ships behind the first "Critical" cyber gate. What it means.

Anthropic’s August report moved misalignment risk from Very Low to Low, citing greater uncertainty after cyber disclosures.

AI agents logged 526 chip-material candidate rows representing 193 formulas. One BC4N deposition recipe passed an automated screen, not lab validation.

OpenAI says Astra may reach its Critical cyber threshold, but key test results remain private. Here is what the disclosure proves—and what it does not.

Meta AI can now use email and calendar context for recurring tasks. Here is what to test before granting primary-account access and trusting its research.

France's competition authority found OpenAI, Google, and Anthropic hold 84% of the AI-agent market and warned of lock-in. What Opinion 26-A-05 means for you.

Coinbase Business can now accept USDC from AI agents over x402. The on-chain record shows a crypto-native payer cohort, not commercial demand.

OpenAI, Meta, NVIDIA/ServiceNow, and Google shipped enterprise agent platforms in a month — and disagree on what the product is. The market-structure read for buyers.

Google shipped Gemini 3.6 Flash, 3.5 Flash-Lite, and a restricted 3.5 Flash Cyber model. Why the gated cybersecurity variant is the real signal in the release.