Anthropic reports a $16.8B revenue quarter

Anthropic tells investors quarterly revenue hit $16.8B

Anthropic told prospective investors on October 14 that third-quarter revenue reached $16.8 billion and it recorded an adjusted profit, according to reporting summarized by The Next Web. That is revenue earned during the quarter, not an annualized run rate.

• Reuters reported second-quarter revenue of about $11.6 billion, putting the quarter-over-quarter increase at about 45%.
• The Financial Times reported the figures, followed by Reuters and Bloomberg reporting the same revenue number.
• Anthropic has not published the results. No profit figure was given, and the reports do not specify which costs the adjusted measure excludes.
• The Financial Times also reported that Anthropic told investors it was cash-flow positive during the quarter.

What it means Frontier AI competition now includes enormous commercial scale, but builders should distinguish reported revenue growth and adjusted profit from published financial statements and fully explained profitability.

Source: thenextweb.com

METR proposes public reviews of labs’ internal agent risks

METR has proposed independent assessments of whether frontier AI developers can control the agents working inside their organizations. The aim is to publish evidence about internal risks, not certify that systems are safe or establish formal oversight.

• Alignment reviews would examine internal incidents and whether training could conceal more serious misalignment.
• Control assessments would investigate whether internal agents could cause severe harm without being detected.
• A third assessment would measure how much AI automation accelerates a developer’s AI research and capability growth.
• METR says it is ready to conduct all three exercises and encourages developers to involve other external groups and publish evidence. The proposal does not establish that labs have agreed to participate.

What it means Public evidence about labs’ internal agents could broaden AI trust assessments beyond model tests, but this proposal is neither a safety guarantee nor an oversight mandate.

Source: www.tokenpost.com

A sandbox ID is not enough to route agent requests

Perplexity published a concrete reliability lesson from SPACE, its sandbox platform: a resumed sandbox can keep its ID while starting a different execution. A stale route can then return old data without failing an ordinary sandbox-ID check.

• Each request carries short-lived, signed proof identifying which execution may handle it and what access is allowed.
• The receiving worker checks that identity against its own runtime record and rejects mismatches before forwarding the request.
• When a receiver installs a new execution identity, SPACE stops old terminal and file-watch streams. Eligible operations can reconnect once and must pass the checks again.
• Perplexity reports that SPACE handles millions of executions daily with an error rate below 0.0002%; it does not isolate this routing method’s contribution to that figure.

What it means Bind requests to the current execution, not just a persistent sandbox ID, to catch stale routes when agent environments pause, resume, or move between machines.

Source: perplexity_ai

AI workflow of the day

How to request deletion from personal-data brokers

Shared by Peter Yang (@petergyang) · View original post ↗

If contact-data sellers keep passing around your details, Peter Yang’s prompt helps you organize deletion and opt-out requests. It asks an agent to find official routes, draft requests, and track what still needs your attention, without promising that brokers will comply.

1. Give the prompt below to Grok @bot or another agent, as Peter Yang suggests.
2. Provide your name, email addresses, and location when asked.
3. Review email drafts before approving sends, then use the agent’s status list to handle pending requests and confirmation links.

Remove my personal and contact info from B2B contact-data sellers and people-search brokers (e.g. Apollo, ZoomInfo, RocketReach, Lusha, ContactOut, Seamless AI, Hunter, Clearbit, Cognism, Kaspr, SignalHire, People Data Labs, LeadIQ, Adapt io, UpLead, plus any others common in my country).

1. Ask me for my name, the emails I use, and where I live, then look up each provider's current official opt-out route on their own site.
2. Write a short deletion and opt-out request that cites the privacy law that applies where I live (e.g. GDPR/UK GDPR, CCPA/CPRA or another US state law, PIPEDA, the Australian Privacy Act). Request deletion of all my data, an end to any sale or sharing, deletion of linked profiles, and written confirmation.
3. Show me the emails as drafts from my own email account. Send them only after I approve.
4. Fill out web forms for providers that don't accept email.
5. Give me a list of what was sent, what's still pending, and any confirmation links I need to click.
6. Offer to repeat this every quarter.