Level : Beginner
This Week in AI #1: AI and AI agent news from October 15 to 22, 2026
By The Info Agent IA editorial team ·
This week’s AI news is all about agents: a new benchmark tests whether they can do real computer science research, and several papers tackle their security. Here are six stories checked at the source between October 15 and 22, 2026, each with what it means for you, followed by the first edition of our model ranking.
An AI agent is a program that does more than answer: it chains actions together (searching, running code, editing a file) to reach a goal. All sources below were accessed on October 22, 2026.
1. FrontierCS 2: agents meet real research
On October 21, a team of 61 researchers posted FrontierCS 2 on arXiv. It is a benchmark (a standardized test) of 48 open-ended computer science R&D tasks across 13 fields, including machine learning systems, databases, security and robotics. Agents work inside real code repositories. Out of 12 agents evaluated, the authors report that the one built on Claude Opus 5.5 performed best, passing 58% of its runs, while no run passed on 16 of the tasks.
What this means for you. Even the strongest agents still fail often on long, open-ended work. If you hand a complex project to an agent, build in human checkpoints along the way.
2. Trail of Bits wants agent-written fixes checked more carefully
In a blog post published on October 21, security firm Trail of Bits introduced two plugins for coding agents. “patching-bugs” reproduces a bug, fixes it and tests the fix. “regression-hunt” looks for anything the change may have broken elsewhere, known as a regression. According to the authors, the tool confirmed 21 regressions among more than 130 changes that had already been merged.
What this means for you. An agent that fixes one bug can quietly create another. If you code with an agent, always ask it to check that what worked before still works after its change.
3. Detecting prompt injection is still hard
Prompt injection means hiding instructions in a web page or document to hijack an agent. An arXiv paper from October 20 shows that an existing detection method, which watches where the model’s attention goes, performs close to random guessing on the AgentDojo benchmark. The authors propose an improved variant, AttnDA, but acknowledge that carrying it over to real agent settings remains difficult.
What this means for you. No automatic filter catches every attack. Avoid letting an agent read unfamiliar pages and then act on your accounts without asking you first. Our explainer on prompt injection covers the habits that help.
4. PerMIT: protecting agent memory
Agent memory lets an agent keep information from one session to the next, which also makes it a target: a planted piece of data can sit there for a long time. Posted on October 20, the PerMIT paper proposes a memory design where only what the user types and values that have been explicitly confirmed are treated as trustworthy. The authors proved two integrity properties with the Lean 4 proof assistant and report no successful attacks on the Trojan Hippo tests.
What this means for you. This is research, not a product yet. But it backs up a simple rule: check from time to time what your assistant remembers about you, and delete anything that looks off.
5. OpenAI Agents SDK 0.24: better recovery after errors
Released on October 15, version 0.24.0 of the OpenAI Agents SDK improves recovery when something goes wrong: completed work is preserved, and the agent can recover from malformed tool arguments generated by the model. Annotations from MCP tools are now exposed as well. The release notes warn that custom error formatters need updating.
What this means for you. If you build with this kit, read the breaking-changes section before upgrading. For everyone else, it is a sign that reliability is now the top priority for agent tooling.
6. Microsoft Agent Framework and goose: two updates
On October 15, Microsoft Agent Framework shipped its Python 1.22.0 release, adding Python 3.15 support, configurable service URLs for Ollama and OpenAI, and stricter rules around tool approval. On October 16, the open-source agent goose moved to version 1.55.0, which among other things fixes compatibility with Claude Sonnet 5.5 and Haiku 5.5.
What this means for you. Both tools tighten control over actions that a human must approve. To compare their licenses and skill level, see our comparison of AI agent frameworks.
The week at a glance
| Story | Date | Source | Type |
|---|---|---|---|
| FrontierCS 2 | October 21 | arXiv | Research |
| Trail of Bits plugins | October 21 | Official blog | Tool |
| Prompt injection detection | October 20 | arXiv | Research |
| PerMIT | October 20 | arXiv | Research |
| OpenAI Agents SDK 0.24.0 | October 15 | GitHub | Release |
| Agent Framework 1.22.0 and goose 1.55.0 | October 15 and 16 | GitHub | Release |
This week’s ranking: first edition
Our model ranking is based on the Epoch AI index, an average of several public tests. This is its first edition, so there are no moves to report yet. As of October 22, 2026, Claude Sonnet 5.5 and Claude Opus 5.5 share first place, ahead of Gemini 4 Argon and GPT-6 Astra.
| Rank | Model | Organization | Index |
|---|---|---|---|
| 1 | Claude Sonnet 5.5 | Anthropic | 95.9 |
| 2 | Claude Opus 5.5 | Anthropic | 95.9 |
| 3 | Gemini 4 Argon | Google DeepMind | 95.3 |
| 4 | GPT-6 Astra | OpenAI | 94.7 |
| 5 | Claude Opus 5 | Anthropic | 92.7 |
The full AI model ranking is updated every week and explains how it is calculated.
Frequently asked questions
Where does the information in this weekly roundup come from?
Every item is checked against its official source: the research paper on arXiv, the publishing organization's own blog post, or the release page on GitHub. Anything we could not open at the source is left out, even if it is widely shared.
Can I trust a paper posted on arXiv?
arXiv hosts preprints, which means papers are not reviewed by other researchers before they go online. The results are the ones the authors report. They are useful as a signal, but they may later be corrected or challenged.
Should I update my agent tools as soon as a new version ships?
Not necessarily. Read the release notes first, especially the section on breaking changes, and try the update on a low-stakes project. Security fixes, on the other hand, are worth applying quickly.
Why doesn't the ranking show any movement this week?
This is the first edition of the site's ranking. Up and down arrows will appear from the next update onward, once there is a previous week to compare with.