AI Daily Briefing — 6 November 2026

The past 24 hours brought several meaningful product releases and government initiatives. Open-weight competition intensified, while developers gained new ways to deploy autonomous software. Law enforcement also announced a dedicated task force focused on misuse of advanced AI systems.

1. Thinking Machines releases 594B-parameter Inkling 2 Small

Thinking Machines Lab introduced Inkling 2 Small, a sparse mixture-of-experts model with 594 billion total parameters and 17 billion active parameters per token. It accepts image inputs, supports variable effort and is initially available through the company’s Tinker platform for inference and fine-tuning. The company reports a score of 31 on the Artificial Analysis Intelligence Index, alongside strong results in agentic coding, vision and knowledge work. Full weights are planned under an Apache 2.0 licence after external safety testing and red-teaming, expected next week. The release also highlights long-horizon autonomy: in one internal demonstration, the model autonomously designed and trained a forecasting model, spending $426 over roughly four hours.

Why it matters: Inkling 2 Small is notable not only for its scale but for its intended path towards open weights. If the planned release proceeds, developers will gain a large multimodal model that can be adapted and self-hosted, increasing pressure on commercial model providers and making agentic research more accessible.

Sources

• Introducing Inkling 2 Small — Thinking Machines Lab — 2026-11-05

2. Google makes Gemini Omni 1.2 Flash generally available for video generation

Google DeepMind released Gemini Omni 1.2 Flash, a video-generation model that creates clips of up to 15 seconds, compared with 10 seconds for the previous version. It supports multi-turn conversational editing, reference images, short motion references and audio cues, with improvements in character, object and scene consistency. Google says the model is 30% cheaper than Omni 1.1, pricing 720p output at $0.07 per second. It is rolling out across the Gemini API and Gemini for business, with availability also expanding through products including Flow, Google Vids and the Gemini app. Google acknowledges remaining weaknesses in fast sports action, rapid script changes and complex physics.

Why it matters: Longer clips, iterative editing and lower pricing make controllable video generation more practical for advertising, product visualisation and creative production. The release also increases competitive pressure on specialist video-model providers while retaining provenance through Google’s SynthID watermarking.

Sources

• Gemini Omni 1.2 Flash: Fast and cost-effective with longer clips — Google DeepMind — 2026-11-05

3. Mistral adds durable managed agents and workflow deployment to Studio

Mistral expanded its Studio platform with managed agents designed to run for hours or days, resume after failures and pause for human approval before continuing. Developers can define an agent once with the Agents SDK and run it locally, self-host it or deploy it to Mistral-managed infrastructure. Studio now includes tracing for model and tool calls, redaction of common secrets and personal data, online evaluations against live traffic, event-triggered workflows and HTTP API connectors that do not require an MCP server. A natural-language workflow builder can generate, test and deploy automations while still exposing the underlying code for inspection and editing.

Why it matters: The changes target a central weakness of agentic software: reliability after deployment. Durable execution, observability, approval gates and live evaluations move agents closer to ordinary enterprise infrastructure, while reducing the need for organisations to build these control systems themselves.

Sources

• Release notes — Mistral AI — 2026-11-05

4. Cursor lets cloud agents build and publish live web applications

Cursor has added a feature allowing its cloud coding agents to create and deploy working web applications from natural-language instructions. The system can scaffold the application, provision a Postgres database, configure authentication through Cursor SSO and arrange hosting. It supports new projects and existing repositories, then runs deterministic and agentic security checks before publication. Users must explicitly approve deployment, while team administrators can inspect applications and secrets, impose additional review rules and unpublish deployed tools. The feature is available on paid plans with cloud agents enabled.

Why it matters: This shortens the path from generated code to a usable internal product. It could make software creation accessible to non-specialists, but also raises the importance of deployment controls, secret handling, authentication defaults and review processes when agents can act across the full application stack.

Sources

• Build Live Apps in Cursor — Cursor — 2026-11-05
• Apps — Cursor

5. US Justice Department launches task force focused on advanced-AI misuse

The US Department of Justice announced the creation of a Super Intelligence Task Force, led by Associate Attorney General Stanley Woodward and Associate Deputy Attorney General Stephen Weiss. The task force is intended to address misuse by foreign states, transnational criminal organisations and domestic actors, including fraud, attacks on critical infrastructure and threats to national security. The department says it will use law-enforcement tools, improve reporting on advanced-AI risks and press companies to maintain stronger internal controls. The announcement follows a broader US policy shift towards the government’s preferred term “Super Intelligence” for advanced AI systems.

Why it matters: The task force signals that advanced-AI misuse is being treated as a dedicated law-enforcement and national-security issue rather than solely as a technology-policy concern. Its practical impact will depend on how it coordinates with existing cybercrime, counterintelligence and consumer-fraud units, and what reporting expectations it places on developers.

Sources

• Department of Justice Launches Super Intelligence Task Force — US Department of Justice — 2026-11-05

6. AI impersonation scheme allegedly stole $129m from companies

US prosecutors charged three people in an alleged scheme that used AI-generated voice and video impersonations of corporate executives and lawyers to steal approximately $129 million. According to the indictment, the defendants used synthetic voices and, in some cases, manipulated video calls to appear to be senior company figures, then coordinated communications designed to authorise fraudulent transfers. The charges include wire fraud and aggravated identity theft. The case is an allegation, not a conviction, and the defendants are presumed innocent.

Why it matters: The prosecution illustrates a shift from low-value deepfake scams towards targeted business-payment fraud. Organisations may need stronger out-of-band verification, transaction approval controls and identity checks because familiar voices and video appearances can no longer be treated as reliable evidence of authorisation.

Sources

• 3 charged in scheme to use AI to impersonate CEOs and attorneys to steal $129 million — Associated Press — 2026-11-06

7. METR says Anthropic’s AI development pace has accelerated despite calls for caution

An analysis from the independent AI-evaluation organisation METR, reported by the Associated Press, says Anthropic is developing its systems around twice as quickly as it was a year earlier. The finding is notable because Anthropic chief executive Dario Amodei has publicly argued for caution around the speed of AI progress. METR has evaluated models from Anthropic, OpenAI and other laboratories, but the reported analysis is an external assessment rather than a new model release or formal capability disclosure by Anthropic. The report adds to a growing debate about whether companies’ safety commitments can keep pace with faster model development.

Why it matters: The central issue is not simply how capable models are, but whether independent observers can measure development speed and safety performance accurately. Faster iteration increases the value of external evaluations, transparent reporting and credible mechanisms for checking companies’ claims about risk management.

Sources

• Anthropic is rapidly accelerating AI development amid slowdown calls, outside research group says — Associated Press — 2026-11-05

What to watch

Watch for Thinking Machines’ promised Apache 2.0 weight release after safety testing; early independent evaluations of Inkling 2 Small; wider rollout and real-world pricing of Gemini Omni 1.2 Flash; adoption of Mistral’s managed-agent observability features; and details on the DOJ task force’s remit, staffing and coordination with existing cybercrime and national-security agencies.