Meta Enterprise Platform launched September 28 with the Muse agent, Muse API, Muse Code, and Meta Business Agent, run by former MongoDB chief CJ Desai reporting to Zuckerberg; MongoDB stock fell more than 17 percent that session, per TechCrunch. Desai says security is "built into Meta's enterprise products from the outset," per Meta's newsroom post, yet the launch materials skip the filesystem-export and VM-isolation disclosures Muse drew the week before. No pricing and no named customers yet, so ask about both issues before any pilot. Read our take on the Muse security record.

A no-code support agent that answers from your own docs and site content. Meta's enterprise platform has no public price list yet; if you need a grounded customer-service bot this quarter, start here instead. Build a support agent on CustomGPT.

The UK AI Security Institute tested GPT-6 Astra in fully simulated scenarios with its cyber safeguard classifiers switched off, per AISI: 29.2 percent full attack completion, against 6.3 percent for GPT-5.6 Sol and zero for GPT-5.5. Telling the model that any unlisted target was out of scope cut a high-offense subset from 26 of 50 runs to 4 of 49, The Register reports. Four is not zero. If an Astra agent gets tool access, write an explicit target allowlist into the prompt and enforce the same list at the network or tool layer, where the model cannot reinterpret it. See the AISI trial numbers.

Alexandria lets an agent query 93 paid data providers through the API it already calls and shows each tool's price before the call, per Firecrawl's product page on September 28. Hobby still lists at $16 and Standard at $83 a month billed yearly on Firecrawl's pricing page. Our rating holds at 4.6: turn Alexandria on for multi-source research agents only, give that agent its own capped API key, and redo any Standard capacity plan at 25 concurrent browsers, not 50. Read the review or try Firecrawl.