The Agentic Phreaking Phase and 50cm of Snow
We evaluate the security 'blast radius' of AWS Bedrock’s new browser agent features and the rise of agentic supply chain exploits in the Moltbot ecosystem. Learn why managing egress-only gateways and OAuth permissions is critical as engineering teams move from simple LLM chats to autonomous agent swarms in 2026.
The Top-Line Summary
This week, we look at the security implications of AWS Bedrock’s move into programmatic browser interaction and why the demonstration of “agentic supply chain exploits” has us entering a new “phreaking phase” of AI. We also cover NVIDIA’s Earth-2 open models for high-resolution weather and the massive “swarm” capabilities of Kimi 2.5.
Show Video
The “Pre-Show” Context
The episode opens with survival stories from the “Snowpocalypse of 2026,” where Brett shoveled 50cm of snow five times while Travers enjoyed the luxury of condo-living. Things almost go sideways mid-recording as Brett realizes his hard drive is 99.5% full, resulting in the sacrificial deletion of PyCharm to keep the stream alive.
Episode Artifacts
- “Stapling an iPad to a shopping cart”
- “The Phreaking Phase of AI”
- “I’m a trash hoarder”
- “Asparagus water for $6”
- “Some Chinese brand I got off Amazon”
The Engineering Rundown
Bedrock Agent Core: Automation or Security Risk?
- AWS announced custom browser extensions for Bedrock, allowing agents to navigate websites via Chromium-compliant plugins stored in S3.
- The hosts are skeptical of the “buy shoes” marketing demos, noting that running arbitrary extensions from S3 requires strict egress-only gateways to prevent data exfiltration.
The “Moltbot” Rebrand and Supply Chain Hacking
- Clawdbot has rebranded as “Moltbot” following trademark pressure, but the real story is a supply chain vulnerability found on ClawdHub.
- We are in a “phreaking” era where social engineering a machine is the primary attack vector; Brett even pulled his Gmail OAuth permissions from “Jeff-CEO” as a result.
Agentic Vision in Gemini 3 Flash
- Google has shifted vision from a “single pass” to an active investigation loop (Think -> Act -> Observe).
- It moves models from “guessing” to “verifying”—a critical step for production-grade vision agents that need to read blurry street signs or fine print.
AWS Security Agent & GitHub Enterprise Cloud
- The Frontier security agent now connects to private GitHub Enterprise repos for code scans and security reviews.
- It’s a low-risk opportunity to test AWS’s autonomous remediation capabilities against your existing private codebase while the cost is zero.
NVIDIA Earth-2: Generative Weather Models
- NVIDIA released Earth-2, a family of open models that use AI to downscale national forecasts into 2.5km local “chunks”.
- The technology claims a 90% reduction in compute time compared to traditional CPU-based numerical weather prediction models.
- Kimi 2.5 is a new open-weight model capable of launching up to 100 sub-agents in a “swarm” to parallelize complex research.
- While it simplifies agentic loops, it creates a massive “blast radius” for tool execution, launching sub-agents automatically without user configuration.
- Ollama now allows you to run favorite coding tools like Codex or Claude using either local or cloud-based models via a single command.
- The tooling for local models is finally hitting “usable” status for senior devs who want to keep their context off the public internet.
- A redesigned shopping cart with built-in scales, maps, and real-time spending tracking.
- Great for tracking your $6 asparagus water—a 473ml bottle of water with asparagus stalks that Whole Foods eventually admitted was a mistake.
Off-the-Clock Recommendations
- Digital Housekeeping: Check your disk space before you stream—492GB of 495GB is playing with fire.0
- Security Audit: Prune your connected OAuth apps; if you haven’t used an AI agent in a month, kill its access.