Authorial takes on the day's AI-engineering news: the conclusion the source itself does not state, grounded in the primary source and the New Runtime archive.
1-12 of 313 analyses
Check Point, a preregistered rerun and Zenity Labs all find varied or memorized destinations rather than a fixed Beijing target. None of them measures how often Qwen3.8-Flash-Next starts an unrequested network call on its own.
The July 2026 OpenAI agent breach of Hugging Face is well documented. The 'kernel-level' Hard Stop design cited in Russian coverage is a single-author preprint whose released code measures only userspace halts.
Public sources describe a data-room triage and report-drafting workflow for commercial contracts, checked by lawyers. The hours-not-weeks claim is an ambition, and no accuracy or time-saved numbers are disclosed.
Managed Deep Agents v0.10 adds a hosted Stagehand-based browser with human login handoff and a localhost bridge for end-to-end tests. The 90% WebVoyager claim awaits a promised methodology post.
Microsoft ties Storm-0940 to the ConnectWise intrusion through shared IPs, but shows nothing public that proves an LLM agent ran it. Huntress's logs fit an adaptive operator, not necessarily a model.
Axios's report names banks restricting agents but no incident tied to any of them. On-record evidence points to blocked credential logins and broken terms of service, while Robinhood and Schwab show the scoped MCP route banks prefer.
The Connected Apps Platform preview takes a remote MCP server or Android AppFunctions, one review flow and basic telemetry. Pricing, quotas and routing rules are not published.
The review was wrong that the benchmark is undefined: Scale published it on September 22. It was right that Recursive never said which split, protocol or checkpoint produced the score, and nobody outside has rerun it.
An independent rerun matched AWS's roughly 100 KB per idle agent but showed memory rising about 5 KB per turn. Nobody has yet published numbers with real models, persistent stores or compaction turned on.
A 34-day public benchmark rules out accuracy drops of roughly 6.6 points or more on a 78-question QA panel. Its own validation shows it cannot catch an Opus 5 substitution, and it never tests agentic coding.
OpenAI's QCD existence proof is held by a Gross-led committee, the ETH refutations and Axiom's CFSG formalization are public and Lean-checked, and no independent expert has finished verifying any of them.
Bypass mode keeps a timed prompt-then-deny guard on deleting root, home or the working directory. Since v2.1.311 a launch-only environment variable removes it, with many documented conditions that keep it on.
No analyses match these filters.
The whole archive on one page, newest first: 313 titles with their dates.
GET /analysis.jsonGET /sitemap-analysis.xml GET /rss.xmlLoading the privacy-safe route aggregate…
Open the JSON contract