OpenAI's revenue run rate nears $50 billion

Following this week's chip-financing reports, the Financial Times reported that OpenAI told investors its annualized revenue was nearing $50 billion at the end of September. That's below the $70 billion in earlier media reports. CNBC confirmed the roughly $50 billion figure and reported Thursday declines in Nvidia, Oracle and CoreWeave shares.

Firmus withdraws its planned IPO

CNBC

Nvidia-backed Australian data-center operator Firmus withdrew its planned IPO, citing market volatility.

Microsoft, Adobe and six IT firms lose access to PERM

Axios

The Labor Department is halting new and pending PERM applications from eight companies, with no end date given. PERM is typically the first major approval step for employer-sponsored green cards, separate from temporary H-1B visas. Immigration lawyers expect the suspension to discourage sponsorship and deter foreign talent.

OpenAI defends the safety-researcher dismissals

The Verge

Following yesterday's letter from researchers, OpenAI says the three dismissals were for violations of sensitive-information policies and denies that raising safety concerns was the reason. The researchers dispute the misconduct claims.

EU panel reviews loss-of-control incidents

European Commission

The Commission's Scientific Panel meets today to present recommendations on frontier-AI safety and security. The panel and AI Office have prepared questions for the developers of models involved in recent loss-of-control incidents.

NOMOS compiles policy into tool-call gates

arXiv

NOMOS compiles written policies into deterministic tool-call gates, with static checks against tool schemas. In the airline benchmark, violations of encoded policy clauses fell from 66.3% to 2.6% among state-changing calls. Costs to legitimate task completion vary by domain.

MXC provides sandboxed code execution

Microsoft / GitHub

Microsoft's MXC provides sandboxed code execution for agents and AI tools, with an implementation you can inspect.

Reducto rebuilds its MCP tools around verifiable operations

AI Engineer

Reducto describes replacing a broad MCP toolset with fewer domain-aware operations and live pipeline snapshots. Explicit confidence scores and error codes help users inspect uncertainty and prevent agents from guessing missing context.

A durable harness executes the agent's plan

AI Engineer

This talk separates model planning from deterministic execution through durable, idempotent workflows. An SRE demonstration compiles a proposed remediation plan into a Conductor workflow, then checks observed state before the next planning round.

Anthropic adds infrastructure defense and OSS scanning

Axios

Anthropic is offering frontier models, on-site engineers and threat research to companies securing critical infrastructure. Its free OSS Scanner will send vulnerability reports to maintainers without human review, so recipients may need to triage inaccurate findings.

SWE-Journey tests coding assistants across user personas

arXiv

SWE-Journey reports average functionality-test pass rates above 75% with simulated software architects and below 25% with simulated non-coders. The benchmark tests clarification and adaptation across multiple turns, rather than a single coding request.

Leaderboard claims need a stated corruption tolerance

arXiv

In replays of Chatbot Arena votes, a few hundred rigged votes made standard confidence intervals certify false rankings. The authors propose certificates that state how many corrupted records a pairwise claim can withstand, even when attackers adapt to published results.