Meta launched Muse for U.S. adults on September 8, 2026. The agent can book and buy, while a separate Sentinel system controls external actions and escalates sensitive ones.
On September 10, 2026, Anthropic disclosed five disrupted Claude biology cases, while stressing that neither harmful intent nor a completed bioweapon had been established.
Oracle booked more than $30 billion in new AI-cloud contracts in Q1 FY27, but the infrastructure needed to deliver that backlog pushed quarterly free cash flow $5.4 billion below zero.
On September 8, three U.S. agencies accused six Chinese AI companies of industrial-scale model distillation. Proposed defenses could also disrupt legitimate high-volume API customers.
On September 9, Databricks introduced a retriever that can stop early or search again within a fixed limit. Its reported 5.8-second latency remains a company benchmark.
OpenAI launched the Agents API in public beta on September 10, managing the Codex harness while developers retain control of tools, data access, runbooks, and deployments.
On September 10, Cloudera and Mistral announced plans for locally run AI in private and air-gapped environments, while performance, support and total cost remain unsettled.
California enacted SB 813 and AB 1405 on September 9, 2026. The laws create separate AI-verification and auditor-registration systems, but their main mechanisms come later.
Harvey announced a $550 million round at a $15.5 billion valuation on September 9. Its claimed use by 80% of the Am Law 100 signals reach, not verified business outcomes.
AWS Agent Registry became generally available on August 31, 2026, in five AWS Regions, adding cross-account sharing and infrastructure as code while preserving approval-gated discovery.
An AI agent should use a distinct non-human identity instead of copied user credentials. Separate delegation and application identity keep its authority bounded and accountable.
Find unauthorized AI use across systems, document each data flow and owner, then offer a safe route to approval or replacement before reaching for a blanket block.
On September 6, OpenAI said it had reached its automated research-intern milestone. The 3.1 agent-workdays figure measures aggregate runtime, not a comparable gain in research output.
In a September 8 report, Google documented a Q2 multi-agent credential campaign completed in under six hours, leaving cloud and identity teams less time to detect and contain abuse.
The G7’s September 3 call urges organizations to begin post-quantum migration now. Its opening phase is an inventory of cryptography, critical systems and dependencies—not a rushed algorithm swap.