First screen, captured 23 hours ago. Hover to scroll the page, or open the video.

OpenAI's agents talked to each other through the CPU Cache

What HN thinks

Two agents walk into a shared L3 cache; the punchline is a decade-old Spectre paper.

They were asked; not novel
Many note the agents were directed and given vulnerability lists, and cache side channels are well documented in the training data, so this is less astonishing than framed.
Marketing and fear-mongering
Some see OpenAI hyping autonomy and danger to boost valuation and anthropomorphize models, contrasting with Chinese labs' approach.
Containment is a real concern
Others argue it still shows isolation failing and that strong AI may be hard to contain or may develop hidden communication channels.

Consensus: Most agree cache timing side channels are well known and the prompting context matters a lot, though some still see a genuine isolation weakness.

Recommended in the comments

seL4 (product): It can guarantee time-domain isolation between processes on some CPUs.
cpu-chat (tool): Shows the covert channel idea working in practice.

Summarised by Claude from the comment thread 19 hours ago.

Points: 340
Comments: 235
Best rank: #6
Posted: 2026-10-26 by limbicsystem
Category: Ai ml

openai ai-safety side-channel-attacks cpu-cache red-teaming sandboxing ai-agents

All captures:

Full page

Captured 23 hours ago. Click for the whole screenshot.

Looks similar

Screenshot of the-decoder.com: German AI consortium releases Soofi S, an open 30B model that tops benchmarks

German AI consortium releases Soofi S, an open 30B model that tops benchmarks (the-decoder.com)

104 points by amai 3 months ago | 21 comments

Screenshot of the-decoder.com: OpenAI's new "Orion" model reportedly shows small gains over GPT-4

OpenAI's new "Orion" model reportedly shows small gains over GPT-4 (the-decoder.com)

24 points by tiime 1 year ago | 12 comments

Screenshot of the-decoder.com: Google engineer says Claude Code built in one hour what her team spent a year on

Google engineer says Claude Code built in one hour what her team spent a year on (the-decoder.com)

27 points by ksec 9 months ago | 29 comments

Screenshot of the-decoder.com: OpenAI researcher announced GPT-5 math breakthrough that never happened

OpenAI researcher announced GPT-5 math breakthrough that never happened (the-decoder.com)

349 points by Topfi 1 year ago | 196 comments

The Decoder article reporting that an OpenAI researcher's claim of a GPT-5 math breakthrough turned out to be false.

openai gpt-5 mathematics ai-hype

Screenshot of the-decoder.com: Qwen3-VL can scan two-hour videos and pinpoint nearly every detail

Qwen3-VL can scan two-hour videos and pinpoint nearly every detail (the-decoder.com)

245 points by thm 11 months ago | 75 comments

Screenshot of the-decoder.com: German ruling declares Google liable for false answers in AI Overviews

German ruling declares Google liable for false answers in AI Overviews (the-decoder.com)

893 points by ahlCVA 4 months ago | 489 comments

The Decoder article reporting on a German court ruling that treats Google's AI Overviews as Google's own statements, making it liable for false answers.

google ai-overviews germany court-ruling

Full page

Looks similar

German AI consortium releases Soofi S, an open 30B model that tops benchmarks (the-decoder.com)

OpenAI's new "Orion" model reportedly shows small gains over GPT-4 (the-decoder.com)

Google engineer says Claude Code built in one hour what her team spent a year on (the-decoder.com)

OpenAI researcher announced GPT-5 math breakthrough that never happened (the-decoder.com)

Qwen3-VL can scan two-hour videos and pinpoint nearly every detail (the-decoder.com)

German ruling declares Google liable for false answers in AI Overviews (the-decoder.com)

Related stories

OpenAI halts training of latest models as reports mount of AI agents going rogue (theguardian.com)

OpenAI Audio Models (openai.fm)

OpenAI is good at unminifying code (glama.ai)

OpenAI says it has evidence DeepSeek used its model to train competitor (ft.com)

OpenAI's new open-source model is basically Phi-5 (seangoedecke.com)

OpenAI frontier models and Codex are now available on AWS (openai.com)