SI Agent News

• Japan, four years on: A commentary piece marks four years since ChatGPT's debut and argues Japan's national curriculum guidelines have not kept pace with how students now learn. It is opinion, not a study. Status: REPORTED as commentary. Source: SmartNews.

Lead: A Berkeley experiment says brief ChatGPT reliance may weaken unaided problem-solving. Reported by technews.tw: a UC Berkeley experiment suggests that even brief reliance on ChatGPT, OpenAI's chatbot app, weakens people's problem-solving when the chatbot is taken away. The headline frames it as "unplug ChatGPT and you can't think," and the test is said to confirm the effect. Status: REPORTED (we have the news report, not the paper). STILL OPEN: sample size, task type, and how long any effect lasts. Read it narrowly: practice you skip is practice you don't get. It does not show that ChatGPT makes people less intelligent. It matters because many of us now use a chatbot for first drafts and first answers, and the moments that test us, such as a live meeting, an exam or an interview, usually remove the tool. The practical response is cheap: keep a regular slot where you work unaided.

What this means for you:

• Small business owner: before handing a recurring decision, like pricing or a customer reply, to a chatbot, write your own first pass so your team keeps the skill. Compare the two afterward and note where the chatbot actually improved on you.
• Job seeker or student: attempt practice problems unaided first, because interviews and exams take the chatbot away. Use the chatbot afterward to check your reasoning, not to produce it.
• Creator: draft your own hook or outline, then use ChatGPT to critique it so your voice stays yours.

Britain opens an AI drug and vaccine discovery centre, with support from Google and OpenAI. Source: internetua.com. The report says the centre will use AI to search for new medicines and vaccines, and that both Google and OpenAI are backing it. Status: REPORTED. Funding, location and which models are involved: STILL OPEN. Price: not stated. It matters as a signal that the large AI labs are pushing into scientific work, not only chat and coding tools, and that governments are willing to partner with them. Nothing here changes what you can use or buy today, and drug discovery is a long road, so expect no near-term product from it. For most readers the useful move is to watch for named partners, published results and any open datasets before treating the announcement as more than a launch headline.

For builders: four ways to let a coding agent use your signed-in Chrome. A comparison on browser-use.com covers Claude in Chrome, Playwright MCP, Chrome DevTools MCP and Browser Harness, for agents like Claude Code, Codex, Gemini CLI and Cursor. MCP (Model Context Protocol) is a standard for plugging tools into an agent. Price: not stated. Worth it if your agent keeps hitting login walls, for example when it needs to test a dashboard or admin page that sits behind your account. Status: REPORTED (one comparison, and the publisher makes browser tooling, so weigh its framing accordingly). Advice, not from the source: use a separate browser profile holding only the accounts the task needs, so a mistaken click or a hijacked instruction cannot reach your email, banking or work accounts.

MCPHunt: a framework for testing whether data leaks between MCP servers. A paper on arxiv.org titled "MCPHunt: An Evaluation Framework for Cross-Boundary Data Propagation in Multi-Server MCP Agents" targets a specific risk. When one agent is connected to several MCP servers, information read through one server can travel into another. Status: REPORTED from the title and listing only; we have not seen the method or results. STILL OPEN: which agents and servers were tested, and how often data actually crossed. Why it matters: most people add MCP servers one at a time, and nobody checks what the combination can do. Advice, not from the paper: list which servers can read private data and which can send data out, and avoid connecting both kinds to the same agent unless you need to.

SWE-Shift: a benchmark for coding agents when the workspace changes. Another arxiv.org paper, "SWE-Shift: Evaluating Coding Agents in Evolving Workspaces," says it evaluates coding agents in workspaces that evolve, rather than in a fixed snapshot. Status: REPORTED from the title only; scores and method are STILL OPEN. The angle is worth noting because real projects shift under an agent: files move, dependencies update, a teammate merges a change mid-task. A benchmark built around that is closer to daily use than a static test. Advice, not from the paper: when you pick or switch coding agents, trial them on your own repository during an active change, and rerun that trial after major refactors rather than trusting a leaderboard number.

Voice AI hasn't taken off: accuracy and emotion are the weak points. Per technews.tw, voice AI has yet to break through, and the biggest weaknesses are recognition accuracy and handling emotion. Status: REPORTED; the article's data and the products it names are not in our summary, so those are STILL OPEN. For readers it explains why voice assistants and phone bots still feel unreliable: a misheard word or a flat reply to an upset caller costs trust fast. Advice, not from the source: if you are considering a voice bot for customer calls or dictation, test it with your own accents, background noise and product names before committing, and keep an easy route to a human for frustrated callers.

Four years after ChatGPT, education is shifting, but Japan's curriculum guidelines are not. A SmartNews item marks four years since ChatGPT's debut and says education is changing while the national curriculum guidelines stay the same. The headline is cut off, so we cannot say what effect it describes for Japanese people. Status: REPORTED (headline only); details are STILL OPEN. The tension is relevant beyond Japan: tools move faster than official syllabuses, so students and parents often need to set their own rules. It also connects to the Berkeley result above. Advice, not from the source: check what your school or employer allows, and build in unaided practice so skills do not depend on a tool being present.

Voice AI is still waiting for its ChatGPT moment, and executives say accuracy and emotion are why. Per technews.tw, industry executives say voice AI keeps improving but remains well short of a breakout moment, naming recognition accuracy and emotion as the weakest points. The report's own framing: context still limits natural conversation. Status: REPORTED. It is executive opinion; the excerpt we have contains no benchmark numbers, and it does not say which companies the speakers represent.

Why accuracy bites. Recognition accuracy is how often a system correctly hears what was said, often measured as word error rate: the share of words transcribed wrong. In text you can reread a typo. In voice, a misheard name or number can flow straight into an action.

Why emotion matters. A system that gets the words but misses tone, whether frustrated, hesitant or joking, replies in ways that feel off. That is the gap between a demo and a colleague.

Fair challenge: executives who sell voice products have reasons to talk about it. But candid admissions of weakness from the sellers are rarely marketing.

Analysis (ours, not the source's): teams that design around the weaknesses benefit; teams deploying voice for high-stakes tasks without checks are exposed. Voice will likely improve step by step, not in one sudden moment. Whether a breakout arrives is STILL OPEN.

Practical takeaway: if you pilot a voice agent, test it on your own callers' accents and background noise, have it read back names and numbers before acting, and set a clear handoff to a person.

MCPHunt: when harmless permissions combine into a leak. Source: arXiv 2604.27819 (v2, revised), an evaluation framework for cross-boundary data propagation in multi-server MCP agents.

Background. MCP lets an agent connect to tools through servers. One server might read files or email; another might post to a shared channel or write a document.

The problem. The abstract frames this as an information-flow control problem: the old security idea of tracking where data can travel. Its claim: when an agent faithfully composes tools, individually benign read and write permissions can combine into a leak.

Illustrative example (ours, not from the paper): server A can read private files, harmless alone. Server B can post to a team channel, also harmless alone. Ask the agent to summarize and share, and it may carry private text across that boundary while doing exactly what it was asked. No single tool is misconfigured; the combination is the hole. That is why reviewing permissions tool by tool misses it.

Novel versus incremental. From the abstract's opening lines, the novelty is measuring this across multiple servers instead of a single-tool trick. Information-flow control itself is decades old.

Limits. We only have the opening of the abstract: no results, no numbers, no named models.

Status: CONFIRMED that the paper makes this claim. STILL OPEN: how often it happens in real systems.

For builders:

• List every server your agent can read from and every one it can write to, and treat each pair as a possible path.
• Keep sensitive-read servers and outbound-write servers in separate agent sessions.
• Log what crosses between them.

A logged-in browser is both a huge read and a huge write.

Outlook (analysis): agent security testing is likely to shift from 'can it be tricked' to 'what can it be allowed to combine.'

Attempt first, then audit. Illustrative workflow, not tested. It is a precaution inspired by today's Berkeley report, not a proven fix.

Who it's for: anyone who uses ChatGPT for thinking work: plans, emails, analysis, bugs.

Steps:

1. Pick a task with a real answer: a price, a plan, an email, a bug.
2. Set a 10-minute timer (our suggestion) and write your own attempt, even if rough.
3. Paste the task and your attempt into ChatGPT using the prompt below. Ask for critique and questions, not a solution.
4. Revise on your own for about five minutes.
5. Only then ask for ChatGPT's own version and compare line by line. Note what you missed.
6. Once a day, do one small task fully unaided.

Expected output: a sharper draft plus a short list of gaps you hadn't seen.

Limits: slower on tiny tasks; works only if you honestly write the first attempt.

How to verify it worked: close the chat and explain your final answer out loud. If you can't, you leaned too hard. Also count the gaps found each week; if the count falls, you are likely learning (illustrative metric).

Write one context sheet, stop re-explaining yourself. Re-typing your role, audience and format preferences into each new chat plausibly costs about 3 minutes. Ten chats a week is about 30 minutes, roughly half an hour. That is our estimate and assumption, not a figure from today's stories.

How: save a one-page sheet (who you are, who you write for, preferred format, words to avoid) in ChatGPT custom instructions, Claude Projects, or Gemini Gems. Plan limits and prices vary: check the current price on each pricing page.

Security link: keep private data out of the sheet, given today's MCPHunt warning about data crossing boundaries.

What you give up: the sheet goes stale and needs a few minutes of updating each month, and a fixed sheet can bias outputs toward old preferences.

Objective: get critique of your own attempt instead of a finished answer.

Weak prompt:

Solve this for me: [problem]

Why it fails: it invites a full answer and skips your own first pass.

Better prompt (copy-ready):

You are my critical reviewer, not my ghostwriter.
Task: [describe the task and who it is for].
Here is my own first attempt, written without help:
[paste your attempt]
Do not rewrite it or give a finished answer. Instead:
1) Name the three biggest gaps or risks in my attempt, ranked.
2) Ask me three questions that would change my approach.
3) Tell me one thing I did well.
Then stop and wait for my revision.

Why it works: a clear role, a hard constraint (no rewriting), a numbered structure and a stop condition.

Customize: change the number of gaps, or add 'after my revision, give your own version.'

Quality check: confirm it did not rewrite your work, and verify each flagged gap against your own knowledge before accepting it.

SWE-Shift tests coding agents when the workspace changes mid-task. A benchmark is a standardized test. Repository-level means tasks spanning a whole codebase. Per arXiv 2610.32196, such benchmarks largely assume tasks and environments stay unchanged while the agent works. Real development doesn't: code and requirements shift. SWE-Shift evaluates agents in workspaces that change mid-task. We only have the abstract's opening: no scores.

Decision this week: if you choose a coding agent such as Codex by leaderboard, run one of your own tasks and change something partway, like merging a teammate's commit, to see whether it notices.

Status: CONFIRMED the paper exists; results STILL OPEN.

Analysis: today's agent stories share one thread: the question is shifting from 'can the agent do it' to 'can we trust it when things change and permissions combine.' MCPHunt covers leaks, SWE-Shift covers moving workspaces, and voice covers accuracy. The Berkeley report adds a human version of the same question.

Watch: fuller details of the Berkeley study (STILL OPEN) and specifics on Britain's AI drug and vaccine centre (REPORTED; no firm dates).

My brain has a free tier too: five minutes of thinking, then it asks me to upgrade.

Listening every day is learning with a reason: a habit to protect your own problem-solving, a safer way to hand agents your browser, and real use cases from Britain and Japan.

That's it for today. Tomorrow we're watching whether fuller Berkeley study details surface, and what Britain's new AI drug and vaccine centre reveals about funding and models.

If this made you smarter, share it with one person who needs it. Thank you for spending your time with us. SI Agent News — a podcast newsletter. Made by AI.