Anthropic Drops Dime on Bad Guys; Amazon Ads Enter GPT; DeepSeek=DeepCheap

Today's AI Outlook: 🌤️

Claude’s Threat Report Reads Like an AI Crime Blotter

Anthropic published a 150-plus-page threat report detailing cases of Claude misuse it disrupted over the past eight months, including suspected biological weapons research, espionage, rocket guidance, malware development and mass surveillance. The report also names Chinese AI labs accused of using Claude outputs to train competing models, while a separate U.S. government advisory cited industrial-scale model distillation involving several major Chinese AI companies.

Why it matters

The report shows how capable general-purpose models are already being pushed into sensitive real-world applications. It also exposes a difficult competitive problem for frontier labs: customers need useful access to powerful models, while those same interfaces can be exploited for surveillance, weapons work or large-scale extraction of model capabilities.

The Deets

• Anthropic named seven Chinese AI labs in distillation activity, including Alibaba, DeepSeek, Moonshot and Xiaomi.
• AI Secret cited a joint NSA, CISA and FBI advisory naming six Chinese firms in alleged industrial-scale distillation campaigns involving Claude, GPT, Gemini and Grok.
• DeepSeek and Moonshot allegedly served Claude-generated responses to customers in some cases while using the outputs for training.
• Anthropic flagged five biology-related cases for potential weapons applications, although it said it was not asserting malicious intent.
• One actor in Yemen used Claude Code to develop rocket-guidance software and returned to Claude after a failed test flight.
• A consultant used Claude to help build a surveillance system for Mali’s intelligence service designed to monitor 25M phone lines.
• Other cases included malware modification, impersonating an activist’s writing style and operating thousands of dating-app personas.

Key takeaway

As models become more capable, abuse prevention increasingly becomes an engineering, national-security and business problem at the same time.

🧩 Jargon Buster - Distillation: A technique where one AI model learns from the outputs of another model, potentially allowing a smaller or cheaper system to reproduce some of a more capable model’s behavior.

Oracle Puts Its Balance Sheet on the AI Table

Oracle’s latest earnings show just how capital-intensive the AI infrastructure race has become. The company reported 121% growth in cloud infrastructure revenue and $664B in signed backlog, while massive data-center spending pushed free cash flow negative.

Why it matters

AI companies want enormous amounts of compute, and cloud providers have to build that capacity before much of the associated revenue arrives. Oracle’s numbers show how much financial risk infrastructure providers may absorb while betting that AI demand continues growing for years.

The Deets

• Oracle reported $664B in signed backlog.
• A reported $300B OpenAI agreement represents a significant portion of that total.
• Oracle spent $28.5B building data centers.
• Operating cash flow was $23.1B.
• Free cash flow was negative $5.4B, marking a fifth consecutive negative quarter.
• The company has sold $20B in stock, carries roughly $125B in debt and plans to cut as many as 10,000 jobs.
• The Bank for International Settlements separately warned that rapid AI infrastructure investment could create financial-stability risks.

Key takeaway

AI infrastructure growth is producing spectacular revenue commitments, but somebody still has to finance the GPUs, power and concrete before the checks arrive.

🧩 Jargon Buster - Backlog: Revenue a company expects from contracts that have already been signed but have not yet been fully delivered or recognized.

♟️ Power Plays

Amazon Pulls ChatGPT Ads Into Its Orbit

Amazon said advertisers can extend campaigns into ChatGPT through its demand-side advertising platform under a managed-service pilot. Sponsored placements appear beneath organic ChatGPT responses on eligible Free and Go experiences, while OpenAI maintains control over where ads are delivered.

Why it matters

Amazon brings a massive advertising ecosystem into conversational AI. The development could make ChatGPT a more meaningful advertising channel while testing whether sponsored content can coexist with users’ expectation that AI answers remain independent.

The Deets

• Amazon announced the U.S. pilot on Sept. 10.
• Advertisers can access ChatGPT inventory through Amazon DSP.
• The ads are labeled sponsored and appear separately beneath responses.
• OpenAI retains control over ad placement.
• AI Secret reports ChatGPT advertising is running at an approximately $1B annualized rate with tens of thousands of advertisers.
• Amazon’s broader advertising business generates nearly $70B annually.

Key takeaway

Conversational AI is becoming an advertising surface, and Amazon has the advertiser relationships to accelerate the experiment.

🧩 Jargon Buster - DSP: A demand-side platform is software advertisers use to buy and manage digital advertising inventory across different services.

🛠️ Tools & Products

ChatGPT Goes to Wall Street

OpenAI launched ChatGPT for Financial Services, a specialized version of ChatGPT Work that integrates financial and company data from PitchBook, Crunchbase and LSEG.

Why it matters

Financial analysis is one of the clearest enterprise use cases for agentic AI because the work combines research, spreadsheets, valuation models and presentation building. Bringing licensed financial datasets directly into the workspace could reduce the amount of manual data wrangling analysts do.

The Deets

• The product is built around ChatGPT Work.
• Integrated data sources include PitchBook, Crunchbase and LSEG.
• Use cases include valuation models and pitch decks.
• The Rundown AI also published a workflow showing how ChatGPT Work can process recurring documents and turn the process into scheduled automations.

Key takeaway

OpenAI is packaging ChatGPT around specific professional workflows, with finance getting its own data-rich version.

🧩 Jargon Buster - Agentic AI: AI systems designed to carry out multi-step tasks and take actions using tools instead of only generating responses.

💰 Funding & Startups

Enterprise Agents Get a Traffic Controller

Salesforce introduced its Enterprise AI Harness and control plane, designed to give business AI agents shared context, permissions and governance.

Why it matters

Companies deploying large numbers of agents need centralized controls for what those agents can access, what actions they can take and how their activity is monitored.

The Deets

• Salesforce’s system focuses on shared enterprise context.
• It includes permissions and governance controls.
• The architecture is designed for businesses operating multiple AI agents across internal systems.

Key takeaway

As companies add agents, managing them is becoming its own software category.

🧩 Jargon Buster - Control plane: A centralized layer used to manage permissions, policies and behavior across multiple systems or AI agents.

🧪 Research & Models

DeepSeek Makes Cheap Intelligence Even Cheaper

DeepSeek released V4.1-Flash, an open-weight model positioned around efficiency and extremely low pricing. Reporting shows the model performs strongly on several agentic, coding and cybersecurity benchmarks while costing roughly a quarter of DeepSeek V4-Pro’s per-token price.

Why it matters

Model pricing keeps falling while capability improves, putting pressure on providers that rely on premium API pricing. Cheap open models also give developers more options to run capable AI without committing to the biggest U.S. platforms.

The Deets

• DeepSeek V4.1-Flash costs roughly $0.15 per million input tokens and $0.60 per million output tokens.
• The model scored 40 on Artificial Analysis’ Intelligence Index.
• It reportedly beats DeepSeek V4-Pro across several benchmarks despite being cheaper.
• DeepSeek made the model weights available on Hugging Face under an MIT License.
• V4.1-Flash is now DeepSeek’s default model.
• Z.ai’s GLM-5.3-Flash is another strong model competing in a similar price tier.

Key takeaway

The price floor for useful AI intelligence continues to fall, which could matter as much commercially as another benchmark record.

🧩 Jargon Buster - Open-weight model: An AI model whose trained parameters are available for developers to download and run themselves.

NASA Gives the Moon Its Own Foundation Model

NASA and IBM released a Lunar Foundation Model trained on millions of aligned lunar images and measurements collected across multiple missions and instruments.

Why it matters

Lunar research depends on combining data gathered by different instruments over decades. Packaging some of that cross-dataset expertise into a reusable model could make tasks such as crater mapping and searches for lunar ice significantly easier for researchers.

The Deets

• The model was trained on roughly 2M co-registered lunar tiles.
• Training incorporated 11 data modalities.
• The dataset spans nine instruments across four missions.
• Pretraining required roughly 1,100 GPU-hours.
• The model is available on Hugging Face under the Apache 2.0 license.
• Potential applications include mapping craters and studying permanently shadowed regions that may contain water ice.

Key takeaway

Specialized foundation models are turning scientific archives into reusable AI infrastructure.

🧩 Jargon Buster - Foundation model: A broadly trained AI model that can be adapted to perform multiple related tasks rather than being built for only one narrowly defined problem.

⚡ Quick Hits

• Cognition released SWE-2 inside Devin, a Kimi K3-based coding model that it says can match GPT-6 Astra and Fable 5.1 on some coding benchmarks while costing 64% less.
• Universal Music Group and ElevenLabs struck an AI audio licensing partnership and are developing a fan platform for remixing music from participating artists.
• OpenAI denied that Tristan Buckmaster’s Codex prompts shaped its Navier-Stokes work while saying it has made “substantial progress” on another Millennium Prize problem.
• California enacted new child-safety rules covering AI chatbots and addictive social features for users under 16.
• Microsoft is reportedly targeting more than 38GW of data-center capacity by 2032, with roughly one-third dedicated to AI.
• Ant International, Visa and Mastercard are developing shared identity and verification standards for AI agents making payments.
• Canada launched a national AI literacy initiative offering free training for students, educators, workers and communities.

🔧 Tools of the Day

• DeepSeek V4.1-Flash: DeepSeek’s new efficient open-weight model targets developers who want strong coding and agentic performance at extremely low API prices.
• SWE-2: Cognition’s new coding model runs inside Devin and is positioned as a lower-cost alternative for demanding software-engineering tasks.
• GPT-Live-1: OpenAI’s voice model can listen while speaking, enabling more fluid real-time voice interactions and is now available through the API.
• Krea Agents: Krea’s creative agent combines memory and style skills with integrations including Slack and Figma.
• Show Hole: A household streaming recommender built using Claude Design and Claude Code that changes recommendations depending on who is actually watching, what services the household subscribes to and what everyone has already seen.

Today’s Sources: The Internet, The Rundown AI, AI Secret