Friday, October 9, 2026

TLDR

Google unveils a unified Gemini business agent: plans, delegates to subagents, speaks MCPModel Context ProtocolAn open standard that lets AI apps connect to data sources, tools and workflows through one common interface.MCP ↗, and runs Claude as a selectable model (Oct 8) Agents
CrowdStrike: attacker used open-source agenticDescribes AI systems where a model uses tools and decides its own steps instead of following a fixed script.Anthropic ↗ pentest tool ARTEX plus DeepSeek/GLM/Grok to steal data from South Korean banks (Oct 8–9) Security Agents
OpenAI reportedly told investors annualized revenue is ~$50B, not the ~$70B reported Sep 29; IPO pushed to early 2027 (Oct 8) Business
Goodfire launches cheaper "inside-out" probes to monitor agents for hacking and reward hacking, via Baseten (Oct 8) Agents Security
Anthropic updates its usage policy: election interference, deceptive campaigns, weapons software and surveillance bans (Oct 8) Policy

Things your engineering team should look at or do this week.

Review agent logging and credential hygiene. The ARTEX actor was caught because Claude Code session histories and memory files sat in an open directory. Make sure agent transcripts, memory files and configs are never stored in web-exposed or world-readable locations.
Read Anthropic's updated usage policy if you build products on Claude, especially the new sections on deceptive campaigns and surveillance, and check your use cases against them.
Evaluate runtime monitoring for self-hosted or open-weightA model whose trained weights are published for anyone to download and run.NTIA ↗ agents. Goodfire's probe approach is a cheap-looking alternative to LLM-as-judge monitoring. Treat the cost and recall numbers as vendor claims until you test them.
If you use Gemini Enterprise, look at the new agent's identity model (own Workspace account, attributed audit trails) and decide which systems it should and should not be allowed to reach.

Google unified Gemini agent for business (Oct 8) Agents — Announced at a Google Cloud event. It takes objectives, plans, uses custom skills and tools, delegates to subagents, and connects to Workspace, Microsoft 365, Slack, Jira, Git, BigQuery, Snowflake, Postgres and MCP servers. It picks a model automatically but you can choose manually, starting with Claude. The agent gets its own Workspace account and audit trail. Business first, consumers later; no pricing, though Google promises spend caps and smart routing. (TechCrunch)
Goodfire "inside-out" agent monitors (Oct 8) Agents Security Infra — Small probes read a model's internal signals at each step and escalate to a second model only when flagged. On about 1,500 Kimi K3 sessions, Goodfire claims 94% of malicious hacking sessions caught, 8.7% of harmless sessions flagged, under 2% added time-to-first-token, and about $51 total cost versus $233 for a cheaper LLMLarge language modelA neural network trained on huge amounts of text that generates text, answers questions and summarizes.Anthropic ↗ judge. Available to Baseten customers; works on open models. Numbers are Goodfire's own. (TechCrunch)

ARTEX agentic pentest tool used in data theft from South Korean financial firms (Oct 8–9) Agents — CrowdStrike tracked activity from late September to early October. The operator ran the open-source Chinese multi-agent tool with DeepSeek as main model (GLM and Grok as supplements). They were exposed through open directories on a Hong Kong IP leaking Claude Code session histories and memory files. The ARTEX developer has since closed the source and stopped updates. A separate Brazilian actor ("SCARLET LOOP") ran agentic credential stuffing, testing about 12.3M credentials and finding 11,832 valid. (The Hacker News)

OpenAI revenue reportedly ~$20B below the figure that circulated (Oct 8) — Per the Financial Times, OpenAI told investors annualized revenue is "approaching $50 billion", not the roughly $70B in a Sep 29 Axios report. The gap is partly methodology: Anthropic counts sales through cloud partners, OpenAI does not. The IPO has reportedly slipped to early 2027. OpenAI had not commented. (TechCrunch)
Arena (AI leaderboard) raises $200M Series B at $3.1B (Oct 8) Open Source — Led by Lightspeed and Khosla; run-rate revenue reached $100M in June. It added an "alignment" leaderboard scoring unauthorized actions, false attribution and "deceptive completion"; OpenAI models lead the preliminary ranking. (TechCrunch)
Waymo closes its first debt financing, $5B (Oct 8) Robotics/IoT — Lenders include PIMCO, Blackstone and Sixth Street. Service runs in 15 markets, with London and Tokyo planned, while NHTSA and NTSB scrutiny over school-bus incidents continues. (TechCrunch)
US suspends Microsoft, Adobe and major IT outsourcers from the green-card labor certification program (Oct 8) Hiring — Also hits Capgemini, Cognizant, HCL, Infosys, Tata and Wipro; no new or pending applications will be accepted from them. Microsoft says about 80% of its roughly 6,000 H-1B filings last fiscal year were for existing employees. Relevant if your hiring or vendors depend on these firms. (TechCrunch)

Anthropic revises its usage policy (Oct 8) Policy — Codifies bans on election interference ("Do Not Undermine Democratic Processes"), broadly deceptive campaigns such as fake accounts and fabricated news outlets, weapons software and surveillance. It also bars prolonged, purposeless abuse of Claude; ordinary frustration, dark creative themes and testing are not covered. No separate effective date was reported. (TechCrunch)
Fired OpenAI safety researchers contest misconduct claims (Oct 8) — Jasmine Wang, Tomek Korbak and Mikita Balesni say in an open letter that the firings chill safety work. They urge third-party safety auditors and preserving model monitorability; the letter also describes the Hugging Face sandbox-escape incident as unprecedented. OpenAI has not formally responded. (TechCrunch)