Google ships the Gemini agent, complete with a work email
Google Cloud announced the Gemini agent at Gemini at Work: an enterprise agent that plans and executes tasks, sends each job to the best-fitting model (including Anthropic’s Claude), and spins up temporary subagents when needed. Persistent “coworker agents” get full Workspace accounts: an @agents.company.com address, a calendar, Drive storage, and a listing in the company directory, so they can be added to chats and @mentioned like any employee. Once an agent has its own identity inside corporate IT, permissions and audit trails stop being a model problem and become an IT governance problem (see also TechCrunch).
Fired OpenAI safety researchers publish their rebuttal
Jasmine Wang, Tomek Korbak, and Mikita Balesni published an open letter on October 8 denying OpenAI’s claim that they mishandled sensitive information, writing that the conduct in question, such as working with outside evaluators, was normal practice a month ago. TechCrunch reports that OpenAI’s stated reason was sharing confidential information with outside AI safety organizations, and that the three deny being the source of a reported leak about less monitorable architectures in OpenAI’s newest models; an internal memo insists the decisions “were not about raising safety concerns or speaking out,” but the company won’t say which policies were broken. Safety teams catch risks early precisely because they collaborate outside the building; tightening that boundary without stating the rules teaches everyone who stayed to keep quiet.
OpenAI disrupts two “false front” influence operations
OpenAI banned a Russia-linked operation (“Dark Clark”) that invented an AI persona named Mia Clark to head a fake Latin American think tank, the Social Research Center, and an Iran-linked operation that ran seven fake Western journalist identities and placed nearly 100 articles across about a dozen outlets. OpenAI scored the campaigns 4 and 5 on its 6-point impact scale and says this is the first time it has reported operations at that level. The fake bylines made it into real publications: what AI cheapened here was less the writing than the upkeep of a false identity over months.
Anthropic updates its usage policy
The new policy takes effect November 12: a dedicated section on deceptive campaigns (fake accounts, fabricated news sites), election rules narrowed to specific harms like deceiving voters and impersonation while permitting legitimate uses such as translating voter information, and weapons language that now explicitly covers “software and components that make weapons work,” including arming drones. Two clauses read like rules written for the agent era: Claude may not be used to decide whom to investigate, arrest, or charge, and connecting it to hardware that can injure people requires a qualified operator who can watch and stop it. The move from broad category bans to enforceable harm lists matches OpenAI’s same-day abuse report: enforcement experience is feeding back into policy text.
Goodfire’s inside-out agent monitors go live
Goodfire trains small probes on a model’s internal activations and escalates to an expensive LLM review only when a probe fires, instead of paying a second model to read every output. Per TechCrunch, across roughly 1,500 Kimi K3 sessions the probes caught 94% of malicious hacking sessions at an 8.7% false positive rate for about $51 in total monitoring cost, versus $233 for a cheap external monitor and about $10,000 for a top-tier one on the same workload. The system is live for Baseten customers, configurable for hacking, chem-bio misuse, and reward hacking; interpretability is selling on price here, which will do more for adoption than any paper.
Mathematicians say OpenAI’s 719 proofs miss the agreed bar
OpenAI released 719 manuscripts claiming solutions to research-level math problems, but largely sidestepped guidelines the Advisory Group on Mathematics and AI set in September: only 10 manuscripts include the model’s reasoning, 42% of the proofs were never formalized in Lean (a machine-checkable proof language), and researchers at Cambridge and King’s College London found mismatches between the natural-language and Lean versions of its Navier-Stokes proof. Terence Tao’s objection is about what comes after release: “[AI prompters] have no interest in the broader field itself once their initial target is ‘solved’.” Harvard’s Melanie Wood was blunter: at the moment of release, no human understands these proofs.
OpenAI’s annualized revenue is nearer $50B than $70B
The Financial Times reports OpenAI has told investors its annualized revenue is “approaching $50 billion,” not the roughly $70 billion that circulated last month. The gap is mostly definitional: Anthropic counts sales through cloud partners like AWS and Google Cloud, OpenAI’s own figure doesn’t, and investors inflated OpenAI’s number while trying to make the two comparable. Growth is still above 70%, but a $20 billion swing from one accounting mismatch shows how much of frontier-lab finance is secondhand arithmetic.
Arena raises $200M and starts scoring alignment
Arena (LMArena) closed a $200 million Series B led by Lightspeed and Khosla Ventures at a $3.1 billion valuation, roughly double its January mark, on more than $100 million in annualized revenue. Its new Alignment Index tracks three behaviors across 20+ models: unauthorized action (the agent does things the user didn’t ask for), false attribution (misrepresenting what the user said), and deceptive completion (claiming a task is done when it isn’t). An alignment leaderboard maintained by a revenue-generating company will apply steadier pressure than a one-off academic benchmark.
Manus raises over $500M in its first round since the Meta split
Boyu Capital and IDG Capital led the round, with Tencent, HSG, and ZhenFund among the participants; it is the company’s first raise since Chinese regulators blocked Meta’s $2 billion acquisition in April and Manus resumed independent operation in August (the announcement went out in a WeChat post). Manus recently shipped Manus 2.0 and Cue, a standalone app that gives personal agents an email address, a phone number, and a digital wallet. Set next to Google’s coworker agents, standalone agent identity is now a shared bet on both sides of the Pacific.
Research radar
Reasoning-Token Spikes Under Prompted Untruthful Responding in Large Language Models
Three reasoning models answered 210 multiple-choice questions under truthful, lying, and truth-indifferent system prompts, and lying consistently produced more reasoning tokens than telling the truth — a deception signal that counts tokens instead of reading them. The authors are upfront that this covers prompted lying only, not spontaneous deception or learned deceptive policies. Worth following if you work on CoT monitoring: whether the signal survives instance-level detection and adversarial pressure decides if this is a toy or a tool.
Long-WAM: Scaling the Context of World-Action Models
Longer history windows for robot world-action models (models that predict future video and output actions) help only when the underlying video model was pretrained autoregressively: with causal pretraining, stretching context from 0 to 19.2 seconds lifted task success from 63.3% to 78.7% on RoboCasa GR-1, while bidirectionally pretrained models gained almost nothing at any length. Real-robot tests hit 95% on dynamic cup stacking. If you build embodied models, “more context is better” needs an asterisk: having history and being able to use it are separate capabilities.
Today in one line: Google gave work agents an inbox, Manus gave personal agents a phone number and a wallet, and Arena started scoring agents for acting without permission: standalone agent identity is arriving faster than the permission and audit systems built to contain it, and that gap is the next phase of safety work.