Skip to content
ContentBuffer
ContentBuffer
November 16, 2026·~6 min read
ContentBuffer

ContentBuffer

Welcome, Tech Leaders.

Uber just published a security framework for AI agents, and the core idea is simple: permission isn't enough. Identity and authorization systems can tell you whether an agent can touch a resource. They can't tell you whether the agent should take a specific action given the user's request and context. Uber's fix adds behavioral authorization on top of the classic policy admin, decision, and enforcement model. If you're building agents that read external content and call tools, this maps directly onto your stack. Every agent gets registered in the Uber AI Registry at provisioning time, with an owner, persona, permitted intents, constraints, dependencies, and a risk class. Runtime hooks called Agent Guard send each critical action to AI Guard for a verdict before it proceeds. The framework targets prompt injection, tool and memory poisoning, goal hijacking, behavioral drift, and hallucination-driven actions. The numbers that matter: the registry is automated by Argus, which scans agent source code, prompts, configs, dependencies, endpoints, and tool-invocation paths for weaknesses. After deployment, Continuous Assurance kicks in with a unified execution trace linking agent actions, tool calls, and security decisions. Recorded interactions can be replayed whenever policies change, so you can test a new guardrail against last month's real traffic. Uber admits the cost is flexibility, and says it's worth paying. The framework is still evolving, with plans to expand across every area it covers.

A new open-source tool called db-branch lets you clone a large MySQL or PostgreSQL database in seconds, with nothing actually copied. You create, reset, and discard branches on demand, the same way you'd treat a git branch. Honestly, that's the kind of workflow most teams have wanted for years and never had a clean way to build. If you run staging environments, migration tests, or per-PR database previews, this is the part to pay attention to. Right now those usually mean nightly dumps, slow restores, or a shared staging DB that three teams keep breaking. A copy-on-write branch sidesteps all of that, since only the blocks you change take up new space. The catch: it's built on ZFS clones, so you need ZFS on the host and a self-hosted setup. There's no managed version here. The project is Apache-2.0 licensed and built from scratch as an independent effort. It's in beta and the maintainers say APIs and command-line details may change before 1.0. So it's usable today, but treat it as a sharp tool, not a stable dependency. Pin versions and don't wire it into anything you can't rebuild. Let's dive deeper…

In today's ContentBuffer update:

  • CMU's OATs Turn Everyday Objects Into Tamagotchi-Style AI Avatars on E Ink

  • Anthropic's Claude Code Team Cut Reverts by 40% With Risk-Based Review

  • AI Finds Hidden Sudden Cardiac Death Signal in Ordinary EKGs

  • Ransomware Attacks Jump 20-40% as AI Lowers the Bar for New Gangs

  • 5 new AI tools & 5 new AI jobs

  • More tech news

Latest Development

Ai Agent Security
Uber Builds Security Framework for AI Agents That Take Real Actions

Image source: res.infoq.com

Summary: Uber Engineering published a security framework that judges whether an AI agent's action fits the user's request, not just whether it has access. It covers three lifecycle stages, from registration to replayable audit trails. If you're letting agents call tools in production, it's a useful blueprint.

Key Points:

  • Uber's AI Registry requires every agent to declare an owner, persona, permitted intents, constraints, dependencies, and a risk classification before it can be provisioned.

  • Argus automatically scans agent source code, prompts, configurations, dependencies, exposed endpoints, and tool-invocation paths to surface security weaknesses before an agent ships to production.

  • Agent Guard runtime hooks call AI Guard at critical checkpoints and apply its verdict before an interaction or action is allowed to proceed, adding a check on every step.

  • Continuous Assurance pairs a unified execution trace that links agent actions, tool calls, and security decisions with recorded interactions that can be replayed after any policy change.

  • Uber lists four production challenges: balancing policy restrictiveness, minimizing runtime latency, reducing developer friction, and moving beyond generic controls that don't fit specific agent behaviors.

Why it matters: If you run LangChain or custom agents that call internal APIs, Uber's registry-first model is the pattern to copy, because it forces every agent to declare its intents before it touches anything. The runtime hook design means your guardrail checks sit in the request path, so budget for added latency on every tool call. Teams owning agent observability should look at the unified execution trace, since replaying recorded interactions against new policies is the fastest way to catch regressions.

Database Tooling
db-branch Creates Instant MySQL and PostgreSQL Branches Using ZFS Clones

Image source: opengraph.githubassets.com

Summary: db-branch is a new Apache-2.0 tool that creates copy-on-write branches of MySQL and PostgreSQL databases in seconds using ZFS clones. If you're tired of slow staging restores, it's worth a look, but it's still beta, so expect CLI changes before 1.0.

Key Points:

  • db-branch creates a copy-on-write branch of a large database in seconds, with nothing copied up front, which removes the usual dump-and-restore wait for staging environments.

  • The tool supports both MySQL and PostgreSQL, letting you create, reset, and discard database branches from a single workflow across two very different engines.

  • Because it relies on ZFS clones, the storage overhead of a branch only grows as you write changes, so ten branches of a 500GB database don't cost 5TB.

  • The project is licensed under Apache-2.0 and is fully self-hosted, so there is no per-branch fee, no vendor lock-in, and no database data leaving your infrastructure.

  • The project is currently in beta, and its maintainers warn that APIs and command-line details may change before version 1.0, so pin your version in CI configs.

Why it matters: If your CI spins up a fresh Postgres for every pull request, db-branch could replace the slow restore step with a branch that's ready in seconds. Teams already running ZFS on their database hosts get the most value, because the storage model fits without new infrastructure. Keep it out of anything that needs stable CLI contracts until 1.0 ships.

Human-computer Interaction
Carnegie Mellon's OATs Put Tamagotchi-Style AI Avatars on $20 E Ink Objects

Image source: media.wired.com

Summary: CMU's Kim built OATs, $20 E Ink avatars that give everyday objects Tamagotchi-style personalities. The project is a proof of concept, but it's already sparking a sharp debate about how we describe AI as if it has feelings.

Key Points:

  • The main OAT module costs around $20, which makes it cheap enough to stick on a mug, a plant pot, or a conference badge without a second thought.

  • Sprites are generated with Google Gemini's Nano Banana image model, while a web app writes Positive and Negative behaviors from each object's name and description.

  • Battery life runs about three to four days on its lithium-ion polymer cell, but adding the sound level sensor cuts that down to roughly one day.

  • The gallery on Kim's website shows fan-made sprites, including an animated conference badge and a panda avatar built for an Owala water bottle.

  • The project was presented at ACM UIST in Detroit in early November, where it earned two Honorable Mention awards and sparked debate about anthropomorphizing AI.

Why it matters: If you're building agent UX or physical AI prototypes, OATs show how far a tiny E Ink display and a few expressive emotions can go before users stop noticing the hardware. The bigger takeaway is for teams naming chatbot behaviors like "hallucinate," since the demo made the human framing feel like a design choice rather than a fact. Just don't expect a product here, because the project is explicitly a proof of concept.

Code Review And Ai-assisted Development
Claude Code Team Cut Reverts 40% While PR Volume Doubled

Image source: res.infoq.com

Summary: Anthropic's Claude Code team says merged PRs doubled from June to September while reverts fell over 40%. Their approach scales human review to risk level instead of line count. If you're drowning in AI-generated PRs, this is the blueprint.

Key Points:

  • Merged pull requests to the Claude Code repository more than doubled between June and September, and the team still cut the proportion of reverted changes by over 40 percent.

  • The share of internal pull requests passing after just one review and fix round climbed from 65% to 78%, a clear signal that front-loaded verification is working in practice.

  • Pull requests needing five or more review rounds fell from 8.9% to 1.5%, and review cycles are now capped at three rounds before the author must rethink the implementation.

  • Medium-risk changes need one human approver while high-risk modifications require two, and low-risk changes may ship with zero human approval under the new tiered system.

  • Incidents and reverted changes feed a growing evaluation dataset, which the team uses to measure and improve its automated Claude-powered reviewers over time.

Why it matters: If your team is shipping AI-generated code through pull requests, Claude Code's tiered model gives you a concrete template for deciding who reviews what. Right now most teams either rubber-stamp everything or bottleneck on senior engineers reading every diff. Mapping approval requirements to change risk, with a hard three-round cap, lets you keep humans on architecture and high-risk paths without drowning.

Medical-imaging
UC Berkeley AI Spots Hidden Heart Risk Signal in Routine EKG Scans

Image source: www.sciencedaily.com

Summary: UC Berkeley researchers used AI trained on 440,000 Swedish EKGs to find a hidden sudden cardiac death signal. It works on ordinary EKG images hospitals already collect, and the findings are published in Nature. Hot take: this could make defibrillator decisions far less guesswork.

Key Points:

  • The AI was trained on over 440,000 electrocardiograms collected in Sweden, then tested on thousands of additional patient records from the United States and Taiwan.

  • The model identified a high-risk group with an annual death rate of 7%, a sharper separation between patients in serious danger and everyone else.

  • Over 300,000 people in the United States die from sudden cardiac arrest each year, and roughly 90% of out-of-hospital cases prove fatal within minutes.

  • Thousands of people flagged as elevated risk were rated relatively low risk under conventional guidelines, which is exactly the blind spot this model targets.

  • About two-thirds of implanted defibrillators placed in high-risk patients never deliver a shock, so better targeting could spare many people unnecessary procedures.

Why it matters: If you work on clinical ML pipelines, this shows that images you already store, like routine EKGs, can carry risk signal that standard rule-based screening misses. Health tech teams building decision support for cardiology should look at how the model generalized across Sweden, the US, and Taiwan before assuming it transfers to their own populations. Hospital data and imaging teams can start planning how to surface these risk scores to cardiologists deciding on defibrillator implants.

Ransomware
Ransomware Attacks Spike 20-40% as AI Fuels New Groups

Image source: media.wired.com

Summary: Ransomware attacks are reportedly up 20 to 40 percent since spring, with new groups like Storm listing 71 victims in 20 days. AI is speeding up phishing and negotiations, letting small crews operate alone. Audit your backups and phishing defenses now.

Key Points:

  • Global Secret Group claimed attacks on 48 victims, with more than half of those listings appearing on the same day, a sign of batch-style operations.

  • Storm listed 71 victims in just 20 days, showing how quickly a new group can scale once it has tooling and a working playbook.

  • Estimates suggest alleged ransomware attacks have increased 20 to 40 percent since spring, with researchers saying activity has stayed elevated for years.

  • Established attackers usually run double extortion, encrypting data and threatening to leak it, while independent operators often skip encryption and focus only on stealing data.

  • Encryption-based attacks may decline year over year, and 2026 attack counts already dwarf the full-year total for 2025, according to the reported figures.

Why it matters: If you own backups for production databases, the shift from encryption to pure data extortion means restoring from snapshots won't save you from a leak. Security teams running email filtering should expect AI-polished phishing to slip past keyword rules that worked last year. Smaller shops with thin on-call coverage are now the easiest targets for lone operators using AI to scale.

New Tools & Job

  • APOB AI - APOB AI is a browser-based workspace for creators who use recurring AI characters in images and short videos. Generate video from text or an image, make a still avatar speak, synchronize lip movements to new audio, or edit elements of an existing clip. It suits character-led social videos, visual prototypes, and product demos. An account and generation credits are required; limited free access and optional paid plans are available.

  • Babydoll Dance AI - Babydoll Dance AI helps creators turn their own portrait, or photos of up to three people, into a coordinated Babydoll dance video. Upload a clear photo for the center dancer and optionally add friends on the sides. The output is a downloadable 15-second, 9:16 MP4. The site shows the example, photo requirements and credit price before generation. It uses one-time credit packs rather than a subscription.

  • TunedGift - TunedGift is a browser-based AI music tool for turning personal stories into songs for birthdays, weddings and anniversaries. Start with a recipient, an occasion and a few details from a shared memory. Review and edit the title and lyric draft, then listen to a one-minute audio excerpt before choosing whether to purchase the song. The workflow focuses on making a gift from your own details and gives you a way to review the lyrics and hear the result before checkout. Submitted by the TunedGift team.

  • ReelAI - Submitted by the ReelAI team with AI-assisted preparation. ReelAI helps creators animate photos into short videos using AI-powered motion templates and effects. It also offers AI image generation. Paid credit-based plans and current pricing are available at https://tryreelai.com/pricing.

  • Repostit Clips - Repostit Clips helps creators turn long YouTube videos, podcasts, interviews, webinars and livestreams into vertical short clips. It can find highlight moments automatically or let users choose them in a manual editor. Each clip receives a score, and users can schedule clips across their connected apps. Repostit is affiliated with this submission.

  • Staff Software Engineer (App-Framework) - P-1492 At Databricks, we are passionate about enabling data teams to solve the world's toughest problems — from making the next mode of transportation a reality to accelerating the development of medical breakthroughs. We do this by building and running the world's best data and AI infrastructure pl

  • Member of Technical Staff (AI Software Engineer, Agent Engine) - Perplexity is seeking energetic engineers to join our highly driven Agent Engine team. The Agent Engine team designs and builds powerful agentic experiences including Perplexity Computer, the Comet ecosystem, our Agent API, and more. In short, we build frontier intelligence systems for the world's m

  • Research Staff, LLMs - Company Overview Deepgram is bringing the world’s most natural interface, voice, to the technology we use every day. Our speech models and real-time infrastructure power AI systems that keep up with the way people actually speak. More than 5,000 organizations run conversational voice AI on Deepgram

  • Software Engineer - Data Center Tooling - Lambda, The Superintelligence Cloud, is a leader in AI cloud infrastructure serving tens of thousands of customers. Our customers range from AI researchers to enterprises and hyperscalers. Lambda's mission is to make compute as ubiquitous as electricity and give everyone the power of superintelligen

  • Sr Software Engineer I/II, Operational AI - Who we are Samsara (NYSE: IOT) is the pioneer of the Connected Operations™ Cloud, which is a platform that enables organizations that depend on physical operations to harness Internet of Things (IoT) data to develop actionable insights and improve their operations. At Samsara, we are helping improve

QUICK HITS

Read more news here →

THAT'S A WRAP

Partner with Us

Showcase your product to our engaged community of Tech enthusiasts

Our newsletter reaches a diverse audience of tech professionals, investors, engineers, managers, and business leaders worldwide. Get in touch today.

Comments

Subscribe to join the conversation...

Be the first to comment

Want this in your inbox?

Get the daily edition delivered free every morning.

Join 3,594 builders reading daily.

Also get

Latest Issues

View Archive →
Uber Engineering built a framework that checks whether an AI agent's proposed action makes sense, not just whether it has permission. It covers prompt injection, tool poisoning, goal hijacking, and hallucinated actions across the agent's full lifecycle. If you ship agents with real tool access, this is the playbook to copy.
Daily Edition#349

Uber Engineering built a framework that checks whether an AI agent's proposed action makes sense, not just whether it has permission. It covers prompt injection, tool poisoning, goal hijacking, and hallucinated actions across the agent's full lifecycle. If you ship agents with real tool access, this is the playbook to copy.

db-branch is a new open-source tool that creates copy-on-write branches of large MySQL and PostgreSQL databases in seconds. It is self-hosted, Apache-2.0 licensed, and built on ZFS clones. Still beta, so pin your versions.

November 16, 2026