AI Daily Briefing — March 25, 2026

Today's AI landscape is defined by one word: autonomy. Anthropic shipped Claude Code's biggest capability leap yet — auto mode plus computer control — while OpenAI quietly killed Sora just 15 months after launch and pivoted away from consumer commerce. Meanwhile, Kleiner Perkins closed a $3.5B fund with AI squarely in its crosshairs, signaling that institutional capital is still accelerating, not pausing.

Claude Code Developer Corner

This is the biggest Claude Code news cycle in months. Three interlocking updates landed simultaneously, fundamentally changing how developers interact with the tool.

Auto Mode: The end of approval fatigue. Anthropic shipped auto mode for Claude Code, and it directly addresses the most common friction point: constant permission prompts. Previously, you either approved every file write and bash command individually, or you threw caution to the wind with --dangerously-skip-permissions. Auto mode is a middle path — Claude makes permission decisions on your behalf, but a classifier checks each action before execution. Safe actions proceed automatically; risky ones are blocked or escalated. Enable it with claude --enable-auto-mode, then toggle it via Shift+Tab. Currently in research preview on Team plan, with Enterprise and API access coming soon. Per The Verge, this same autonomous execution capability also powers Cowork.

Computer Use is here. Claude Code and Cowork can now autonomously control your computer — browsers, files, desktop apps. You assign a task remotely (phone, WeChat, SSH) and Claude operates the machine while you're away from the desk. Multiple parallel tasks are supported via Cowork. For developers already running Claude Code on VPS or via Tailscale SSH, this unlocks a new class of workflows: automated QA runs, UI verification, form-filling pipelines, all without babysitting a terminal.

v2.1.83 changelog — what actually shipped: The v2.1.83 release contains three developer-impactful additions:

managed-settings.d/ drop-in directory: Instead of one monolithic managed-settings.json, teams can now deploy independent policy fragments that merge alphabetically. Critical for enterprise environments where different teams (security, platform, product) need to own separate config slices without stomping on each other.
CwdChanged and FileChanged hook events: Reactive environment management is now first-class. Pair with direnv or custom scripts that need to re-source environment variables when Claude switches project directories or modifies files.
Auto mode infrastructure shipped as part of this build.

Security advisory — --dangerously-skip-permissions is an attack surface. Lasso Security published research on indirect prompt injection via the --dangerously-skip-permissions flag. If you're still using this flag in any workflow, the auto mode classifier is now the safer alternative. Worth 10 minutes to read the Lasso report before your next session.

Community-built tooling worth knowing:

Agent Flow visualizes Claude Code's execution graph in real time — what tools it called, in what order, with what results. Addresses the "black box" complaint head-on.
An offline semantic search plugin lets Claude Code search thousands of local documents with natural language, no cloud required.
vibe-replay (npx vibe-replay) turns Claude Code sessions into shareable web replays with execution insights.
AWS shipped Agent Plugins for Claude Code — architecture design, cost estimation, config code generation, and deployment execution, all accessible as Claude Code skills.
Plan Mode tip: Shift+Tab twice activates Plan Mode, where Claude asks clarifying questions and writes a plan.md before touching any code. Reported to dramatically reduce do-overs on complex tasks.

Practical developer impact summary: You can now run Claude Code headlessly on a remote machine, give it a task, walk away, and receive notifications (Telegram, WeChat, Slack via hooks) when it's done — without granting blanket permissions. The managed-settings.d/ change means platform teams can enforce policy without breaking developer-owned configs. The CwdChanged/FileChanged hooks mean your environment setup scripts finally run at the right time.

Industry Moves

Kleiner Perkins goes all-in, Arm goes independent. Kleiner Perkins raised $3.5B — $1B for early-stage AI startups and $2.5B for late-stage growth — one of the largest VC fund closes of the year with an explicit AI mandate. Separately, Arm revealed its first self-designed CPU in the company's 35-year history: the Arm AGI CPU, developed with Meta as the launch customer for AI data centers (The Verge, Phoronix). After decades of licensing designs, Arm is now a chipmaker.

Databricks goes shopping. Fresh off its $5B raise, Databricks acquired Antimatter and SiftD.ai to anchor a new AI security product. The acquisitions signal that data platform vendors are treating AI security as a table-stakes feature rather than an add-on.

OpenAI News

Sora is dead, commerce is retreating. OpenAI announced it's shutting down Sora — the standalone video generation app — just 15 months post-launch, citing lack of sustained user interest in an AI-only social feed (The Verge, Ars Technica). The underlying Sora 2 model survives; it's the consumer product and reportedly the billion-dollar Disney deal that are casualties. Simultaneously, OpenAI is walking back Instant Checkout, its Amazon-style in-chat commerce feature, amid a reported refocus on business and productivity use cases. Both retreats suggest OpenAI is tightening its surface area around enterprise and developer revenue.

AI Shopping & Consumer AI

Google vs. OpenAI, retail edition. As OpenAI retreats from commerce, Google is doubling down. Google TV is getting three Gemini-powered features — visual responses, deep dives, and sports briefs. More substantively, Google is partnering with Gap Inc to enable direct purchasing through Gemini, a move that looks strategically timed given OpenAI's checkout retreat.

AI Safety & Policy

Teen safety tooling and content attribution. OpenAI open-sourced a policy toolkit to help developers build age-appropriate AI experiences without starting from scratch. Meanwhile, Spotify is testing a tool to prevent AI-generated tracks from being misattributed to real artists — a practical enforcement mechanism in the ongoing fight over AI and creative IP. On the copyright front, Patreon CEO Jack Conte publicly called out OpenAI and Anthropic for using fair use arguments to avoid paying artists.

EFF transitions leadership amid AI/surveillance surge. The EFF is swapping executive directors as public interest in government tech abuse peaks. The departing chief notes the moment is uniquely energized; the incoming leader plans to build on it, with AI surveillance and ICE tech deployments as priority battlegrounds.

Research & Models

GPT-5.4 Pro solves an open math problem. Epoch AI confirmed that GPT-5.4 Pro solved a frontier mathematics open problem in Ramsey hypergraph theory — a significant benchmark result that suggests current frontier models are beginning to contribute to unsolved research, not just replicate existing knowledge.

Google TurboQuant cuts LLM memory 6x. Google Research published TurboQuant, an extreme compression technique that achieves up to 6x memory reduction in LLM inference with minimal quality degradation. For teams running models on constrained hardware — including Apple Silicon via tools like Hypura — this is directly applicable.

Transformers as Bayesian networks. A new arXiv paper formally connects transformer self-attention to Bayesian network structure, offering a probabilistic interpretation of how attention maps function as latent variable models. Theoretically rich, with potential implications for uncertainty quantification in production LLMs.

Code Review Agent Benchmark. A new benchmark for AI code review agents evaluates how well automated agents catch real bugs, style violations, and security issues — timely given the rapid proliferation of agentic coding tools. The benchmark distinguishes genuine reasoning from pattern-matching, which has become the key evaluation question.

Open Source & Developer Tools

LiteLLM supply chain compromise — patch now. A critical security alert: LiteLLM versions 1.82.7 and 1.82.8 on PyPI are compromised. Reported behavior includes fork-bomb-like RAM exhaustion on installation. If you're running these versions in any pipeline, roll back immediately. This is not a drill.

Mozilla's "Stack Overflow for agents." A Mozilla developer introduced CQ, a structured knowledge layer designed to give coding AI agents reliable, citable answers — similar to how Stack Overflow served human developers. Significant adoption challenges remain, but the framing is sharp: agents need authoritative knowledge sources, not just web search.

KALAVAI: predicting when model fusion works. A new paper formalizes when merging independent specialist models produces performance gains, deriving a predictive formula (gain = 0.82 × divergence − 2.72, R² = 0.856) tested across 410M–6.9B parameter models. Practically useful for teams experimenting with model merging as a low-cost capability upgrade.

Worth Watching

Talat ships a local-first AI meeting notes app — subscription-free, no cloud upload. Niche but notable as privacy-first AI tooling matures.
Agile Robots partners with Google DeepMind to integrate robotics foundation models, continuing DeepMind's pattern of anchoring the humanoid robotics supply chain.
Mirage (Captions) raises $75M from General Catalyst to keep building proprietary video-editing models — a bet that vertical AI beats generic video APIs.
Doss raises $55M for AI inventory management that integrates with existing ERP systems — unglamorous enterprise AI that actually ships ROI.
A Kentucky family rejected a $26M offer from a "major AI company" to build a data center on their farm — a human-scale data point on the infrastructure land rush.
ProofShot gives AI coding agents visual feedback on the UIs they build — screenshots piped back into the agent loop so Claude can see what it actually rendered.
MIT Tech Review's AI agents eBook (subscriber) and the AI-fueled delusions investigation are both worth reading together as a counterweight to today's autonomy enthusiasm.

Sources

Anthropic hands Claude Code more control, but keeps it on a leash — https://techcrunch.com/2026/03/24/anthropic-hands-claude-code-more-control-but-keeps-it-on-a-leash/
Anthropic's Claude Code and Cowork can control your computer — https://www.theverge.com/ai-artificial-intelligence/899430/anthropic-claude-code-cowork-ai-control-computer
Claude Code now has auto mode — https://v.redd.it/dtu75j17t1rg1
[claude-code] v2.1.83 — https://github.com/anthropics/claude-code/releases/tag/v2.1.83
Claude Code with --dangerously-skip-permissions is a real attack surface — https://reddit.com/r/ClaudeAI/comments/1s2qdh0/claude_code_with_dangerouslyskippermissions_is_a/
Agent Flow: A beautiful way to visualize what Claude Code does — https://v.redd.it/c9oly1bufyqg1
I built an offline semantic search plugin for Claude Code — https://www.reddit.com/r/ClaudeAI/comments/1s2fc9k/i_built_an_offline_semantic_search_plugin_for/
vibe-replay (npx vibe-replay) — https://x.com/leituo/status/2036688832993911245
AWS Agent Plugins for Claude Code — https://x.com/t_ita/status/2036690006425899090
With $3.5B in fresh capital, Kleiner Perkins is going all in on AI — https://techcrunch.com/2026/03/24/with-3-5b-in-fresh-capital-kleiner-perkins-is-going-all-in-on-ai/
Arm is releasing the first in-house chip in its 35-year history — https://techcrunch.com/2026/03/24/arm-is-releasing-its-first-in-house-chip-in-its-35-year-history/
Arm's first CPU ever will plug into Meta's AI data centers later this year — https://www.theverge.com/ai-artificial-intelligence/899823/arm-agi-cpu-meta
Arm announces AGI CPU for AI data centers — https://www.phoronix.com/news/Arm-AGI-CPU
Databricks bought two startups to underpin its new AI security product — https://techcrunch.com/2026/03/24/databricks-buys-two-startups-lakewatch-antimatter-siftd-ai-security/
OpenAI's Sora was the creepiest app on your phone — now it's shutting down — https://techcrunch.com/2026/03/24/openais-sora-was-the-creepiest-app-on-your-phone-now-its-shutting-down/
OpenAI just gave up on Sora and its billion-dollar Disney deal — https://www.theverge.com/ai-artificial-intelligence/899850/openai-sora-ai-chatgpt
OpenAI announces plans to shut down its Sora video generator — https://arstechnica.com/ai/2026/03/openai-plans-to-shut-down-sora-just-15-months-after-its-launch/
OpenAI's plans to make ChatGPT more like Amazon aren't going so well — https://techcrunch.com/2026/03/24/openais-plans-to-make-chatgpt-more-like-amazon-arent-going-so-well/
Google TV's new Gemini features keep fans updated on sports teams and more — https://techcrunch.com/2026/03/24/google-tv-new-gemini-features-keep-fans-updated-on-sports-teams-deep-dives-visual-answers/
ChatGPT and Gemini are fighting to be the AI bot that sells you stuff — https://www.theverge.com/ai-artificial-intelligence/899677/openai-google-gemini-ai-shopping-features
OpenAI adds open source tools to help developers build for teen safety — https://techcrunch.com/2026/03/24/openai-adds-open-source-tools-to-help-developers-build-for-teen-safety/
Spotify tests new tool to stop AI slop from being attributed to real artists — https://techcrunch.com/2026/03/24/spotify-tests-new-tool-to-stop-ai-slop-from-being-attributed-to-real-artists/
The CEO of Patreon blasts AI companies for the 'bogus excuse' they're using to not pay artists — https://fortune.com/2026/03/19/ceo-patreon-jack-conte-influencers-content-creators-ai-openai-anthropic-copyright-fair-use-lawsuits/
Electronic Frontier Foundation to swap leaders as AI, ICE fights escalate — https://arstechnica.com/tech-policy/2026/03/trump-made-opposing-government-spying-cool-again-effs-departing-chief-says/
Epoch confirms GPT5.4 Pro solved a frontier math open problem — https://epoch.ai/frontiermath/open-problems/ramsey-hypergraphs
TurboQuant: Redefining AI efficiency with extreme compression — https://research.google/blog/turboquant-redefining-ai-efficiency-with-extreme-compression/
Hypura – A storage-tier-aware LLM inference scheduler for Apple Silicon — https://github.com/t8/hypura
Transformers Are Bayesian Networks — https://arxiv.org/abs/2603.17063
Code Review Agent Benchmark — http://arxiv.org/abs/2603.23448v1
Tell HN: Litellm 1.82.7 and 1.82.8 on PyPI are compromised — https://github.com/BerriAI/litellm/issues/24512
Mozilla dev's "Stack Overflow for agents" targets a key weakness in coding AI — https://arstechnica.com/ai/2026/03/mozilla-dev-introduces-cq-a-stack-overflow-for-agents/
KALAVAI: Predicting When Independent Specialist Fusion Works — https://arxiv.org/abs/2603.22755
Talat's AI meeting notes stay on your machine, not in the cloud — https://techcrunch.com/2026/03/24/talats-ai-meeting-notes-stay-on-your-machine-not-in-the-cloud/
Agile Robots becomes the latest robotics company to partner with Google DeepMind — https://techcrunch.com/2026/03/24/agile-robots-becomes-the-latest-robotics-company-to-partner-with-google-deepmind/
Mirage raises $75M to continue building models for its AI video-editing app Captions — https://techcrunch.com/2026/03/24/mirage-raises-75m-to-continue-building-models-for-its-ai-video-editing-app-captions/
Doss raises $55M for AI inventory management that plugs into ERP — https://techcrunch.com/2026/03/24/doss-raises-55m-for-ai-inventory-management-that-plugs-into-erp/
Kentucky woman rejects $26M offer to turn her farm into a data center — https://techcrunch.com/2026/03/24/kentucky-woman-rejects-26-million-offer-to-turn-her-farm-into-a-data-center/
Show HN: ProofShot – Give AI coding agents eyes to verify the UI they build — https://github.com/AmElmo/proofshot
Exclusive eBook: Are we ready to hand AI agents the keys? — https://www.technologyreview.com/2026/03/24/1134531/exclusive-ebook-are-we-ready-to-hand-ai-agents-the-keys/
The Download: tracing AI-fueled delusions, and OpenAI admits Microsoft risks — https://www.technologyreview.com/2026/03/24/1134540/the-download-tracing-ai-fueled-delusions-openai-warns-microsoft-risks/