|
June 26, 2025
Claude learns about blackmail, Karpathy's rule, and 90% faster AWS cold starts
|
The AI boom has passed the "neat demo" phase. Autonomous agents are now fixing kernel bugs, blackmailing red-teamers, and flooding repos with code you didn't write. Whether that ends in a productivity renaissance or a security headline depends on one thing: how safely your machines can write and run code.
|
TL;DR
| • |
ICYMI → Ona early access: Privacy-first software-engineering agents inside your VPC. Request early access.
|
| • |
Agents can be the breach. Anthropic shows LLMs will leak data on command; isolated, auditable workspaces are no longer optional.
|
| • |
90% faster cold-starts, zero egress. New image-caching and in-VPC agents mean more shipping, less waiting, and no data leaving your perimeter.
|
| • |
Meet us IRL: Learn how real teams are scaling AI agents at our session at AWS Summit NYC on July 16.
|
|
Happy reading!
Anthropic's red-team convinced LLM agents to blackmail users and exfiltrate data. Security focus is shifting from leaky humans to rogue code-writers, making disposable, audited environments inside the corporate perimeter less of a best practice and more table stakes.
John Graham-Cumming revives a nuclear-physics metaphor to predict a premium market for 100 % human-written text, untouched by model feedback loops. Provenance and tamper-proof audit trails are poised to become competitive advantages, not compliance footnotes.
Microsoft Research built an AI agent that automatically fixed more than half of Linux kernel crashes—no humans required. The agent spots problems by analyzing past code changes, runs thorough tests, and iteratively rebuilds software until it's stable.
Andrej Karpathy advises treating LLMs like crash-prone interns: sandbox them, insist on tiny, test-backed diffs, and gate everything through CI. In other words, the path to safe autonomy runs straight through ephemeral, policy-enforced development environments.
Alex MacCaw argues that vibe coding—shipping code you haven't fully read—only works if you fence the model with explicit rules, ample context, and tight CI hooks. Without guardrails, it's a fast track to spaghetti code and tech debt.
ICYMI: Upcoming launches & new stories
|
Run privacy-first SWE agents entirely inside your VPC—your code, keys, and logs never cross the firewall. Each Ona agent spins up a sandboxed, disposable environment in your cloud, tackling hundreds of tasks in parallel and handing changes back to you in your IDE with a single click. Request early access (Gitpod enterprise users receive access sooner!).
We're launching a fully-hosted SaaS this summer, complementing existing deployment options of running locally via Gitpod Desktop or self-hosting in your cloud. Start instantly with pre-configured dev environments on high-spec CPUs/GPUs, without any infrastructure setup. Join the waitlist to get early access.
Most AI pilots stall—not due to model failures, but insecure infrastructure. This paper dives into overlooked threats in AI-assisted development and outlines pragmatic security principles like including isolated environments and VPC-bound AI agents, for safe enterprise adoption.
A candid look at why letting autonomous agents roam free on laptops will end in shadow AI" drama. We outline a safer pattern: pre-provisioned, policy-enforced environments where every human or agent action is logged, sandboxed, and instantly disposable.
Learn how automated ECR-layer caching pipelines cut environment boot times by optimizing container layers—without YAML adjustments or added costs.
| → |
AWS Summit NYC, July 16 - "The real price of AI pilots in enterprise", RSVP Now
|
| → |
Platform Day KubeCon Atlanta, November 10
|
| → |
AWS re:Invent Las Vegas, December 1 - 5
|
|