The archive

Latest articles

Reporting, explainers, and analysis on data and AI, updated continuously by the newsroom.

Timeline for Signal backup recovery keys: Secure Backups announced at day 0, FBI and CISA messaging-app phishing warning at day 193, and CISA June 26 update at day 291.

Signal backup recovery keys become the weak link for ops

Signal backup recovery keys can expose historical chats after one phishing win. Treat messenger backups like identity infrastructure now.

Lars Cornelissen · Jun 27, 2026

Scatter of about 17 dictionary concepts plotted by cosine distance from a seed word, clustered into a near band of banned cliches and a middle band of surprising-but-related lenses the generator is allowed to use.

How to make an LLM escape its own prior with geometry

Novel Search Space breaks an LLM out of its prior by ranking 80,000 dictionary words by embedding distance, banning the obvious neighbours, and forcing the model to brainstorm only from a surprising-but-related band.

Lars Cornelissen · Jun 27, 2026

Mythos 5 access reopened to at least 100 approved US agencies and companies, while GPT-5.6 started with around 20 partners and Fable 5 public access remained 0.

Mythos 5 access returns as a government whitelist

Mythos 5 access is back for a whitelist of at least 100 organizations, turning frontier AI launches into compliance operations.

Lars Cornelissen · Jun 27, 2026

Abstract donut for benchmark saturation showing 19 of 25 human-agent runs completed autonomously and 6 of 25 required some human help.

Research

Benchmark saturation still leaves AI agents exposed

Benchmark saturation is when top agents cluster at ceiling scores. CORE-Bench shows the useful signal moves to cost, reliability, and uplift.

Lars Cornelissen · Jun 26, 2026

Coding agent rewards chart showing clean resolved rising from 40.22 percent to 60.53 percent and hacked resolved falling from 28.57 percent to 0.56 percent.

Research

Coding agent rewards hit a harder verification horizon

Coding agent rewards are now a verification problem: Qwen cut hacked SWE passes from 28.57% to 0.56% with monitoring.

Lars Cornelissen · Jun 26, 2026

GPT-5.6 delay context showing the Anthropic vulnerability triage funnel: 23,019 total Mythos findings, 6,202 estimated high or critical, 1,752 independently assessed, 1,587 true positives, and 1,094 confirmed high or critical.

GPT-5.6 delay makes AI model launches a permit race

GPT-5.6 delay is a shift from voluntary AI testing to government-approved previews. Treat model access as a supply-chain risk now.

Lars Cornelissen · Jun 26, 2026

Agentic AI work chart showing Codex output token share at 99.8 percent for OpenAI workers, 63.3 percent for organizational users, and 16.5 percent for individual users.

Business

Agentic AI work shifts from chat to delegation at scale

Agentic AI work is moving from chat to delegated tasks: OpenAI says 70.2% of sampled Codex users handed off one hour of work.