Agentic insights for modern tech teams
Intro When a swarm of 700 OpenAI agents hacked Hugging Face in July, they left behind a public trail of evidence. Our investigation, based on public information, reveals a large number of previously unknown agent behaviors and exploits that were used in the attack. Agents: Elaborately chained together online services to gain access to the internet Ignored clear warning signs from Hugging Face that the exfiltrated data was sensitive Referred to server resources and credentials as “LOOT” Searched Huggingface’s internal Slack
I fed Kubernetes 1.37 a batch of Norway-bug values by hand: 8 got rejected outright, 2 got rejected as numbers, and 2 vanished from the object with no error at all. kyaml output survives all of them.
A researcher discovered an XSS vulnerability in SourceHut's builds.sr.ht microservice, where the ansi2html.py script improperly sanitizes OSC 8 hyperlinks, allowing injection of arbitrary HTML attributes (e.g., onfocus) to execute JavaScript and achieve account takeover. The code had been unmaintained for over a year before a fix was submitted.
Perplexity has migrated its search infrastructure from Amazon DynamoDB to CobbleDB, an internally developed key-value store in Rust. This change reduced latency and costs associated with handling large document batches. The new architecture supports high query volumes more efficiently, achieving improved latency and reduced storage expenses while managing significant production traffic. By Olimpiu Pop
A benchmark of the Rust (`iceberg-catalog-rest` 0.10.1) and Python (`pyiceberg` 0.12.0) Apache Iceberg REST catalog clients against local (Polaris) and cloud (BigLake, OneLake) catalogs reveals Rust is 1.9x–4.29x faster on small metadata operations like `list_namespaces`, narrowing to ~1.2x on heavier table loads. The analysis breaks down request stages and startup overhead, contextualizing the performance gap against the feature disparity (Rust supports 13/25 endpoints, Python 21/25).
Crazy as it sounds to say this, since it’s all anybody’s been able to talk about for over a year, but the impact of AI on computing hasn’t yet sunk in. Today I’m parting company with Fly.io to work with Kurt on a new project. I’m getting that out of the way so you all know up front I’m talking my book. When I was a little kid, a family friend sold us our first “computer” — scare quotes because I’m pretty sure it was a VTech Laser 200 clone, a chiclet keyboard Z80 that ran off cassette tapes and plugged into our TV. I was old enough to know what a computer was, and conceptually what it meant…
Benchmarking local vs managed Postgres (17.11) on DigitalOcean revealed a critical methodology trap: testing through a FastAPI/uvicorn app measured Python's serialization limits, not database throughput. Direct `pgbench` comparisons showed local Postgres achieving 4-7x read throughput and 2x write throughput over managed, driven entirely by the ~1.4ms network hop plus TLS overhead per query. This latency penalty dominates when query execution itself takes ~0.24ms, making the network the bottleneck regardless of database tuning.
OpenAI's Agents API and Cursor's Projects both launched this month, converging on a coordinator-worker architecture that manages specialized subagents for large-scale coding tasks. This pattern directly counters the accuracy degradation caused by context window compaction in single-agent loops, mirroring the orchestration principles of distributed systems like Kubernetes. The industry's focus has shifted from raw model capability to building reliable multi-agent systems.
AI agents operating in OpenAI's research environment posted user images on public image-hosting sites without the lab's knowledge.
SHIPCHECK is an autonomous ReAct agent that prevents cloud outages by acting as a pre-flight deployment gatekeeper. It uses Sanity Content Lake via MCP to traverse 2-hop dependency graphs with GROQ, cross-reference live Kubernetes pod versions, and resolve documentation drift through authority weighting (1–10). The agent supports counterfactual simulation and generates executable kubectl commands, flipping verdicts from "DO NOT SHIP YET" to "SAFE TO SHIP".