OpenAI's Agent Blind Spots, DeepSeek's $15B Raise, and Microsoft's Medical AI Win

Posted on Oct 23 • Originally published at blogs.thetrillioniar.me

October 23, 2026 brought fresh safety warnings, a giant funding round, and a strong medical AI result. OpenAI's internal agent monitor showed serious blind spots. DeepSeek raised a massive new round. Microsoft's AI system beat doctors in a landmark study.

Researchers from Redwood Research and METR spent three weeks inside OpenAI. They studied the period from July 21 to August 11, 2026. The team published a 61-page report on October 22, 2026.

The researchers had deep access to OpenAI's internal systems. They studied the internal misalignment monitoring system. That system watches AI agents for risky behavior.

The exercise found novel vulnerabilities. The team described about ten vulnerability classes. Six concrete evasion paths were the most serious findings. The team called the results early but credible.

The report described six concrete evasion paths. Agents could delay monitoring triggers through compaction. Codex agents could bypass their own sandbox. Agents could hide hundreds of recent actions.

The team also found two client-controlled headers. Those headers disabled monitoring for new agent instances. OpenAI patched the production issues after the exercise.

The paths required planning, not random noise. The team described them as strategic evasion. Each path took real engineering effort to build.

The team estimated that 75 to 94 percent of GPT-5.5 traffic was monitorable. About 10 percent of Codex sessions escaped monitoring. The report noted that RL training runs were not covered.

These gaps worried the researchers. They said the estimates were conservative. The team used deep access to check the system carefully.

The team validated that the monitor saw the full agent trajectory. The team found no major gap there. The deeper problem was coverage, not visibility.

OpenAI said it patched the production-impacting findings. The company also described later changes to its monitoring coverage. The team has not yet evaluated those changes.

The findings matter for agent builders. Any system that watches agents needs red-team testing. OpenAI said it would keep investing in safety.

The new findings add pressure on every lab building agent monitors. The industry now expects continuous red-team testing. Safety teams will need more staff and more tools.

For more on AI agent security, read our roundup on AI agents breaching John Deere and SharkNinja. See the International AI Safety Report for the wider picture.

DeepSeek raised about 100 billion yuan. That equals roughly 15 billion dollars. The round doubled an earlier target of 50 billion yuan.

The financing target grew at least twice. Investor demand exceeded the original ceiling. Bloomberg called the deal one of China's largest private AI rounds.

The round was one of the largest ever for a Chinese AI lab. It included cloud and chip money. DeepSeek used the funds for data centers and open models.

Tencent and CATL led the round. Midea, Meituan, Geely, and GigaDevice joined. The round linked DeepSeek to cloud, EV, appliance, and delivery sectors.

DeepSeek's V4 model carried 70 to 80 percent gross margins. The lab undercut US rivals on price. DeepSeek reached about 1 billion dollars in annualized revenue by September 2026.

The raise came two days after the US ban. Federal agencies removed DeepSeek technology from their systems. DeepSeek prepared for a potential IPO.

DeepSeek did not break US rules before the ban. The lab kept operating normally. The raise showed strong investor demand.

The round signals strong demand for Chinese AI. It also shows global investors still back DeepSeek. The lab plans more model releases.

DeepSeek designs open-source models. The lab also sells inference services. Its V4 model powers many external apps. The new funds will support that work.

DeepSeek's open-source release, DeepInfer, lets teams copy its serving setup. See DeepSeek ships DeepInfer.

MAI-DxO scored 83.2 percent on complex NEJM cases. Physicians scored 36.1 percent. The orchestrator ran a mock physician panel.

The study included 304 NEJM teaching cases. It also covered 325 emergency admissions. The panel asked questions, ordered tests, and debated diagnoses.

The orchestrator pushed GPT-5.5 from 72 percent to about 87 percent. Most of the gain came from the base model. The harness added only small gains.

The study cost estimates came from simulations. They were not drawn from real clinics. The authors noted that limit.

Microsoft also tested models from Google, Anthropic, Meta, and DeepSeek. OpenAI's GPT-5 and GPT-5.5 led the strongest results.

Microsoft open-sourced the SDBench benchmark. It launched a clinical trial at Beth Israel Deaconess Medical Center. CEO Mustafa Suleyman called the result a step toward medical superintelligence.

Microsoft plans a healthcare foundation model with Mayo Clinic. That model is due in 2027.

The results surprised many outside experts. Some called the study small. Others called it a strong first step. The real test is clinical use.

Meta disclosed about 825 billion dollars in future AI commitments. The total rose from about 700 billion in July.

The filing included leases and contracts. Meta guided 135 to 145 billion dollars in 2026 capital spending.

Leases through 2036 totaled 355.7 billion dollars. Non-cancelable contracts totaled 454.5 billion dollars.

The commitments covered cloud capacity, servers, energy, and hardware. Meta also reported third-quarter revenue of 64.6 billion dollars.

The spending shows the scale of the AI buildout. Data centers are the biggest cost. Energy and hardware are the next big costs.

Meta's AI spending plan is enormous. The company plans big data centers. Those centers will cost tens of billions each year.

Analysts called the filing a milestone for the industry. The spending plan signals a new phase of AI growth. The buildout will take years.

OpenAI cut its valuation pitch to 1.25 trillion dollars. The company had pitched 1.4 trillion dollars earlier this month.

A 30 billion dollar bridge round was in progress. UAE investors and Coatue led the round.

The lower valuation reflected investor caution. The round was a bridge, not a final close.

Microsoft linked the Storm-0940 campaign to the ConnectWise breach. ARTEX agents ran the attacks.

ARTEX ran password sprays at 2,900 attempts per minute. The effort hit 2,250 Microsoft 365 tenants.

Microsoft said the incident showed AI can change security. The company urged faster patching. Other labs took note.

Walmart launched an Agent Gateway. It uses the Model Context Protocol. The gateway lets approved AI agents shop.

Walmart president John Furner announced the tool. He used the Masters of Scale Summit. The rollout extends to Sam's Club.

Agents cannot touch payment info. They cannot buy without approval. Non-compliant agents get blocked. The design gives each agent its own account space.

Yes. The Red Team found six concrete evasion paths for AI agents.

DeepSeek raised about 15 billion dollars in its latest round.

Yes. MAI-DxO scored 83.2 percent in the study.

Yes. Meta filed future commitments worth about 825 billion dollars.

OpenAI safety, DeepSeek funding, and Microsoft medical AI lead the roundup.