Long form
Deep Dives
The longer read: reporting that follows the evidence past the first turn of the story.

Invest
Anthropic's Q3 dollar gain was 21% to 24% smaller than Q2's, and a $2 trillion valuation is 27x a $74bn run rate but 18x $110bn
Anthropic told investors Q3 revenue was $16.8bn, up 45%, but the dollar gain was 21% to 24% smaller than in Q2, depending on whose Q2 figure you use. Annualized revenue estimates run from $74bn to $110bn by year-end, so a $2 trillion IPO valuation is 27 times one and 18 times the other.

Invest
Character.AI files for bankruptcy years after $2.7 billion Google deal, with Disney's $20 million stalking-horse bid
Character.AI filed for Chapter 11 with a $20 million opening bid from Disney, about 0.7% of Google's $2.7 billion deal of 2024. Compute ran $7.0 million a month in 2026 against a best revenue month of $5.3 million, so the bid is the floor to assume for a license-and-hire shell.

Security
UTA0576 targeted Active Directory secrets that the Ivanti patch does not reset
Ivanti patched two of three Connect Secure zero-days on October 13, ten days after Volexity saw UTA0576 start stealing Active Directory credentials with them. Volexity found no implant, so the lasting exposure is the credentials it copied.
Leadership
OpenAI says a physical sweep found the hidden recorder before its models traced it
FBI agents arrested OpenAI engineer Abdulrahman Alfozan on Oct. 9, nine days after a security sweep found a hidden recorder in an executive conference room. Prosecutors cite internal-tool queries and a spouse-linked chatbot account as planning evidence, and no source says either raised an alert first.

Leadership
Google and IAB Europe listed Discover-Tech as certified until Citizen Lab showed them its spyware-linked ads
Citizen Lab says Israeli ad firm Discover-Tech's ads helped install Candiru spyware on at least 45 Al Jazeera devices in 2024 and 2025. Google and IAB Europe listed it as certified until shown the findings, so their vendor lists did not catch it.
Product
Eight states sue OpenAI over the distance between its safety claims and what ChatGPT does
Eight state attorneys general sued OpenAI on Wednesday, each in its own court, alleging it sold ChatGPT as safe while knowing the risks. The complaints set each safety claim against what the company knew. No claim has been tested in court yet.

Product
A third of failed agent runs in AgentMail's test ended at one hidden default
Block, Notion and Refactoring each tell teams to design MCP tools around whole jobs. We'd name one owner for the server and one for its credential scope, because agents in AgentMail's test failed where tools left things unsaid.
Security
GitLab and npm attacks that skip credential theft both start with a repository the attacker creates
GitLab patched CVE-2026-97229, a CVSS 10.0 code-execution flaw, on October 12. Rapid7 found it needs an attacker-planted CI/CD file, while Bloom Security found 1,500+ npm packages open to takeover through freed GitHub names.

Invest
CoreWeave's June bond has gone from yielding 9.625% to 15%
CoreWeave's bonds yielded about 15% on Oct 14 as hot US inflation data helped erase some $80 billion of AI market value. The 2032 bonds sit nearly nine points above the day's 5.63% peak 30-year Treasury yield, so the price reflects CoreWeave's credit as well as the Fed.

Build
Notion and Block push MCP tools toward workflows as the protocol drops its session
MCP's final 2026-07-28 specification, released July 28, 2026, removes the protocol-level session and the initialize handshake. Servers that need state now mint a handle for the model to pass back, and Notion and Block both favor workflow-shaped tools.
Security
Cisco, Palo Alto and Nextcloud app flaws are under attack; Cisco warns upgrading a compromised virtual appliance destroys forensic data
CISA gave federal civilian agencies until October 17, 2026 to prioritize fixes for exploited Cisco and Palo Alto Networks flaws. A patch does not show whether anyone got in first, and a clean result from Cisco's own compromise check is not proof, The Hacker News reported.
Science
Training on one bad habit changed how models answered unrelated prompts
Researchers who fine-tuned GPT-4o on insecure code saw misaligned answers to unrelated prompts in roughly 20% of cases, and about 50% with GPT-4.1. Backdoored variants, including a Terminator one keyed to the year 1984, show the shift can stay hidden until a trigger appears.