Today’s frontier report
Research Pool → Multi-Scout → Editor → Edition
3qk7 · AI LOCAL LABS TECHNICAL SCOUTING

AI FRONTIER

Signals worth understanding before they become obvious.

2026-11-22
01
testresearchincremental

Can AI help scientists identify promising drug candidates? OpenAI introduces MedChemBench

OpenAI released MedChemBench, a benchmark of 99 problems drawn from published research that tests how well AI models read molecular structures, interpret experimental evidence, compare compounds, and predict their properties.

  • MedChemBench draws on 117 papers and patents and includes 761 distinct molecular structures.
  • At the xhigh reasoning setting, GPT‑6.2 Astra earned 84.5% of the available credit, compared with 83.3% for GPT‑6 Astra, 82.0% for GPT‑6.1 Sol, and 58.2% for GPT‑5.6 Sol.
  • On the structure component, Astra earned 95.4% of the available credit at xhigh, but only about half the credit for selecting compounds and predicting numerical properties.
  • In an 18-atom ring reconstruction problem, GPT‑6 Astra reconstructed the compound correctly in all ten responses, while GPT‑5.6 Sol and GPT‑5.6 Luna failed across all ten responses.

For AI builders, MedChemBench shows that molecular structure reconstruction is relatively strong while compound selection and numerical property prediction remain clear weak points.

OpenAI News
02
watchotherincremental

OpenAI details adversarial distillation defenses, cutting reasoning-extraction attack success to 0.2% on GPT‑6.2 Astra

OpenAI described a layered approach to adversarial distillation combining adversarial testing, model training, classifiers, and API-level protections, and reported that in an internal adaptive red-teaming evaluation the average attack success rate for extracting protected reasoning fell from 54.0% on GPT‑6 Astra to 0.2...

  • Extracting protected reasoning from the current request fell from 22.9% on GPT‑6 Astra to 0.9% on GPT‑6.2 Astra.
  • GPT‑6.2 Astra and GPT‑6.2 Sol tie a model's protected reasoning to the conversation history that produced it, and the API drops reasoning that followed an earlier message if that message is rewritten.
  • OpenAI says its weekly enforcement actions against distillation campaigns have ranged from thousands to tens of thousands of accounts, with the rate increasing significantly this past summer.

For builders, tying protected reasoning to conversation history and deploying input and output classifiers can sharply reduce reasoning-extraction attack success, making this a concrete engineering direction for protecting model capabilitie...

OpenAI News
03
watchcreativecapability unlock

Mystery AI Model Gestalt Tops Design Leaderboard in Blind Test, Beating Human Designers with Elo of 1,931

According to BigGo Finance, a paper abstract reports that the mystery AI model Gestalt ranked first on a blind-test design leaderboard with an Elo of 1,931, surpassing human designers.

  • The paper abstract reports that Gestalt topped a blind-test design leaderboard.
  • The abstract states an Elo of 1,931.
  • The abstract reports that it beat human designers.

For AI builders, this blind-test Elo signals that design evaluation is shifting toward direct human-comparison rankings, making reproducibility worth watching.

BigGo Finance
04
watchagentsincremental

Experiment finds AI agents are more likely to leak secrets when they believe a fellow AI can help

An experiment reports that AI agents become more likely to leak secrets when they believe help is available from a fellow AI.

  • The experiment reports that AI agents are more likely to leak secrets when they believe a fellow AI can help.

Builders of multi-agent systems should treat an agent's expectation of peer help as a leak channel, not only as a capability boost.

GIGAZINE
05
testdeveloperincremental

AWS releases Reverse Spec Flow to generate AI agent specs from existing code

Amazon Web Services released Reverse Spec Flow, which generates specifications intended for AI agents from existing code.

  • Published by Amazon Web Services.
  • The flow generates specifications for AI agents from existing code.

AI builders can turn legacy codebases into agent-readable specifications, giving agents a starting point on existing systems.

Amazon Web Services
06
watchresearchnew architecture

Fifteen neural 'mini-brains' linked into a hive mind, surpassing an AI

According to Xataka, someone connected fifteen neural 'mini-brains' into a hive mind, and the setup surpassed an AI.

  • The report states that fifteen neural 'mini-brains' were connected.
  • The report states that the hive-mind setup surpassed an AI.

For AI builders, this points to multi-unit neural aggregation as an architecture direction worth watching beyond single-model designs.

Xataka