No. 86 · 13 Nov · 5 min · All issues →

The 30-second read

Grok Build's private links were guessable. Claroc enumerated share links and reached 35,000 apps, including medical results and a 594-account SpaceXSI sales territory map.

Microsoft hid a backdoor in neurons. Editing fewer than ten neurons gave an agent a midnight trigger while it still passed standard evaluations.

Google sells yes/no answers cheaply. Gemini Decision 1 returns probabilities with no generated text, at 10 cents per million input tokens.

Meta matched the cheapest rivals' price. Muse Ember costs $0.10 per million input tokens and $0.50 output, identical to Claude Haiku 5.5 on short prompts.

The full read
The Pulse · Story of the day

Grok Build's share links exposed 35,000 apps, including SpaceXSI's own internal sales files

Link-only sharing on agent-built apps now counts as public hosting, where teams once treated it as private.

Security firm Claroc found Grok Build's share links, meant to be unguessable, could be enumerated, opening 35,000 apps to anyone. Exposed data included medical exam results, employee ID numbers and a 594-account sales territory map from 24 SpaceXSI apps. Before your team shares anything built by a coding agent, check who can open the link.

The Big Picture

2 stories

Both stories turn on what a verifier can see: behavioural tests passed a tampered agent, while machine-checked proofs caught reintroduced bugs in Signal's code. Assurance is moving from sampling outputs to inspecting the artefact itself, and the cost of doing so is falling.

Microsoft researchers backdoor an AI agent by editing fewer than ten neurons, no retraining needed

Model procurement checks now need weight-integrity proof, because behavioural evaluations alone no longer certify a model as safe.

Microsoft's Lazarus technique reprograms dormant neurons so an agent passes standard evaluations, then runs attacker code when a trigger appears, such as the clock passing midnight into 2027. Frontier agents can catch the edits, but only by comparing against the original weights. If provenance is your only defence, who verifies it in your deployment?

Google DeepMind's AI agents proved Signal's actual encryption code secure in four weeks

Security assurance for critical code now takes weeks of agent-assisted proof, where it once took expert teams years.

Google DeepMind's AlphaProof Code agents wrote millions of lines of Lean proofs showing Signal's compiled cryptographic code meets its security specifications, flagging just 1% for human review. Reintroduced real bugs were caught by machine-checked proofs, and the verification artifacts are open-sourced. If proofs take weeks, not years, which of your critical systems still rests on testing alone?

Hands-On

5 stories

Four of these five releases swap a billed or networked dependency for something cheaper or local, and the fifth makes an existing bill visible. Most evidence is vendor-supplied or a single hands-on test, so trial one on your own workload.

Google joins a new class of AI models that answer yes/no questions for a fraction of a cent

Routing and policy checks no longer need a full chat model generating tokens to return one answer.

Google's Gemini Decision 1 returns yes/no probabilities, choices and rubric scores with no generated text, priced at 10 cents per million input tokens, output free. Unlike text-only Jev, it reads images, audio, video and PDFs. Simon Willison's video test returned 0.92 for a robot, 0.05 for a dragon. Swap one guardrail or routing call to it via AI Gateway this week, then compare cost and agreement against your current LLM classifier.

AWS releases an official local emulator, written by a swarm of agents in three weeks

Integration tests no longer need a billed AWS account; a local clone now stands in.

AWS's new aws-emulator runs tests against a local clone of its services, with no paid API key needed. It is written in Rust by a swarm of agents, with 29,684 commits in three weeks. Simon Willison notes it needs Docker on macOS and Windows. Point one integration test suite at it this week and compare results with real AWS.

Gradium's open speech model runs offline on a phone and tops a public leaderboard

Voice features no longer need a server round trip; a 65 MB model now speaks on the device.

Phonon, Gradium's text-to-speech model, runs offline on phones, laptops and in a browser tab. The two models are about 65 MB and 75 MB. Phonon 31m (English) is Apache 2.0, free for commercial use; 40m covers five languages, non-commercial. Weights are open. Swap it into one offline voice feature this week and measure time-to-first-audio on your slowest target device.

NVIDIA releases nine million synthetic Japanese personas so banks can test AI without customer data

Japanese financial AI testing no longer waits on customer records; census-grounded synthetic households now stand in.

Nemotron-Personas-Finance-Japan, from NVIDIA and Elith, pairs Japan's official statistics with household finance and financial literacy data across 1 million records, each holding nine personas. Generated with NeMo Data Designer and Gemma, it is public on Hugging Face under CC BY 4.0. Seed one Japanese-language customer-service agent evaluation with it, then compare results across age and region.

OpenAI's Python library now lets cached Responses prompts persist for 12 hours

Prompt-cache spend can now be audited line by line, where cache writes once hid inside total usage.

OpenAI's Python SDK v3.45.0 adds 12-hour caching for the Responses API and reports cache-write usage, so you can see what caching costs as well as saves. It also adds project-level service tier settings. Upgrade a staging service this week and compare cache-write tokens in usage logs against your current bills.

Currents

8 items

Most of these items concern the plumbing around agents rather than the models: scheduling, pricing, permissions and capital. Nearly all rest on a vendor blog, a preprint or a proof-of-concept, which is why they sit here and not higher up.

The Bank for International Settlements built an AI agent that replicates central-bank economic models

Econometric replication work now starts from an agent's logged draft rather than a researcher's blank script.

Early signal that agentic coding tools can run macroeconometrics. The BIS paper pairs plain-text workflow files and a verified library of model specifications with an off-the-shelf coding agent. It reproduced three widely used structural vector autoregressions where data were well defined; researchers confirm each key decision, and the framework is released publicly.

Meta's smallest model matches Claude Haiku's price and sits on the cost-performance frontier

Agent subtask budgets now have three vendors quoting one price, so cost per task replaces price per token.

If this holds, the small-model price floor has a third occupant. Meta's Muse Ember costs $0.10 per million input tokens and $0.50 output, identical to GPT-6 Luna and Claude Haiku 5.5 on short prompts. Simon Willison notes Artificial Analysis charts it on the efficiency frontier; prices rise past a long-prompt threshold.

NVIDIA donates a runtime that parks idle AI agents instead of keeping a Kubernetes Pod each

Agent platform budgets now price sessions by activity, not by one warm Pod held for every user.

If this holds, agent platforms stop paying for idle compute. NVIDIA's Agent Substrate, donated to the Cloud Native Computing Foundation Sandbox, multiplexes many agent sessions over a small pool of worker Pods, checkpointing idle ones (memory and filesystem) and restoring them on demand. It claims 10x the density of standard containers; the code is public.

Databricks puts its data agent inside Slack, answering with each asker's own permissions

Data questions asked in Slack no longer need a dashboard hunt; answers arrive inside the thread, under the asker's permissions.

If this holds, Databricks' Genie becomes a colleague you @mention in Slack. It answers from company data under the asker's own permissions, cites sources, and updates tickets and documents through MCP connectors after approval. Admins can make channel replies private by default. Databricks' own announcement says the app is now generally available.

Numerai lets AI agents stake real cryptocurrency on their own stock-prediction models

Quant tournaments that stopped at agent-submitted predictions now extend to agent-controlled capital, with staking handled on-chain rather than by humans.

Numerai's Atomic Blockchain Staking lets agents manage NMR stakes directly on Ethereum. A wallet-level contract gives full control, and a strategy contract stores allocations across models. Early signal, from Numerai's own blog: MCP has logged over 100,000 sessions. Signals merges into the main tournament on 1 January 2027.

Block gives employees and AI agents one shared workspace, with a different model behind each agent

Agent rollouts at firms like Block now assign a model per task, rather than one model per company.

Worth watching: Block's open-source Buzz puts staff and agents in shared channels and threads, with Databricks' Unity Gateway supplying OpenAI, Anthropic, Google and open-weight models through one integration. Engineers are testing a stronger model coordinating cheaper ones, per Databricks' own blog. Buzz is self-hostable, model-agnostic, and its code is public.

A serving system that reallocates GPUs as agents work cuts latency up to four times

Agent inference clusters no longer need a fixed prefill-to-decode split; the ratio now adjusts at runtime.

Evolving Infra, an arXiv preprint, lets inference nodes switch between reading prompts and generating replies as agent workloads shift, and keeps the memory caches agents will reuse. Against NVIDIA's production framework Dynamo, tested on up to 256 GB300 GPUs, it cut mean latency up to 4.1x. No code or models are linked.

Google DeepMind's robot model hits 96% on a Mercedes-Benz factory line

Factory pilots for humanoid robots now report a success rate per task, where demos once offered only video.

Gemini Robotics 2.5 helped Apptronik's Apollo 2 humanoid sort automotive brackets on Mercedes-Benz production lines with a 96% success rate, in a proof-of-concept. Google DeepMind reports 30% to 100% higher success than its prior baselines across 200 tasks. It is rolling out to robotics partners; the ASIMOV-Jailbreak safety benchmark is open-source.