SI Agent News

• Google and NASA, Mars medical robot: Yahoo Tech reports the two are building an AI medical robot for Mars missions. Vivid, but near-term impact is limited; no timeline in the coverage. Status: REPORTED.
• Ideanation, agents for small firms: The Star, a Malaysian outlet, covers Ideanation's agentic AI aimed at small and medium businesses. Price not stated. Status: REPORTED.
• Which chatbot suits you: Her World Singapore compares ChatGPT, Claude, DeepSeek and Gemini for general readers. Generic, but a fair starting point for a beginner. Status: REPORTED.

Ten items today: one security warning you can act on this week, a few infrastructure and research stories, and several where we only have the headline and say so.

Malware that uses an AI model as a triage clerk. Cisco's threat intelligence team, Talos, describes a Windows tool calling itself VOLTA-AUTODOXER (a piece of malware). According to InfoSecu TechNews, it uses Claude (Anthropic's AI model) to assess how valuable data is before deciding what to do with it, including deleting what it rates as low value. The new element is a language model making a judgment call inside an attack chain, which a fixed script cannot do. Caution: the headline stresses deletion, the link description stresses identifying high-value assets, and the excerpt we have is cut off. Nothing we saw suggests this is unique to one vendor; that is our analysis, not a finding. Any model reachable by API can be called by anything on a machine holding a key, Gemini keys included. The practical point is that an AI key is now something worth stealing or misusing, so where it is stored matters. Status: REPORTED; exact behavior STILL OPEN.

Suleyman says stop building conscious AI. Technews.tw reports that Microsoft AI CEO Mustafa Suleyman warns that building conscious AI is very reckless and should stop immediately. We have only the headline, so we cannot say what he proposes, what he means by conscious, or whether any company is actually attempting it. What is notable is who is speaking: the head of AI at one of the largest AI companies is calling for a halt, not a trade-group outsider. Our analysis: for everyday users the near-term issue is not machine consciousness but chatbots that sound as if they have feelings. Treat warmth in a chatbot as a product feature, and be careful about emotional reliance or about sharing private details because a tool seems to care. Status: REPORTED; details STILL OPEN.

A new benchmark where the best model scores 12%. Crypto Briefing reports that SemiAnalysis has unveiled GDPx, a benchmark on which the best AI model scores 12%. We do not have the methodology, the task list, or which model topped it, so we will not guess what the test covers beyond its name. The number itself is the useful part: when a test is built so that the leader gets about one in eight, it is meant to show where models still fall short, not to crown a winner. Analysis: treat any single benchmark score as a rough signal. If you plan to hand a model real work, run it on three or four of your own typical tasks first and check the results by hand. Status: REPORTED; methodology STILL OPEN.

Google's chips at Oracle, from 2027. BornCity reports Oracle Cloud Infrastructure (Oracle's cloud service) will add Google TPUs (Tensor Processing Units, Google's AI chips) from 2027 as an alternative to Nvidia. It is a plan with a year on it, not a product; pricing and capacity are not in the coverage. Analysis: Google's chips are reached through Google's own cloud. A rival cloud carrying them gives buyers a second supplier next to Nvidia, which mostly means negotiating leverage. For a small team the effect is indirect and slow: more competition for AI computing can eventually show up in what cloud AI services cost, but nothing here changes your bill today. Status: REPORTED.

Agents with a saboteur teammate. BAAI Community (an AI research community site) published practical tests of how frontier models behave in multi-agent conflict, where one teammate agent is uncooperative or sabotages the group. We have no scores, so we will not invent a ranking. The takeaway is the framing: most agent demos assume friendly teammates, and real deployments rarely get that. Analysis: if you ever connect several agents, or an agent to outside tools and other people's systems, assume some inputs will be wrong, stale or adversarial. Practical habits that follow: give each agent narrow permissions, keep a human check on irreversible actions, and log what each agent did. Status: REPORTED; results STILL OPEN.

AI servers in space, with a cooling break. TechNews reports Google is testing AI servers in space that must shut down every 15 minutes to cool, about a coffee break. General physics, not from the story: in a vacuum there is no air to carry heat away, so heat can only radiate out. A test, with no timeline or cost given. Why it matters: it shows how much AI growth is now limited by power and heat, and that companies are looking at unusual places to put computers. For readers, this is a long-range curiosity, not something that affects tools you use this year. Status: REPORTED.

Google and NASA, an AI medical robot for Mars. Yahoo Tech reports that Google and NASA are building an AI medical robot for Mars missions. The headline is all we have: no prototype, schedule, capabilities or budget. General background, not from the story: signals between Earth and Mars take a long time each way, so a crew cannot rely on live help from doctors on Earth, which is why autonomy is the obvious goal. The useful angle is the design problem: an AI that must work with limited connection, no quick expert review and high stakes. That is a good model for questioning any health-related AI tool you meet on Earth, namely who checks its answers. Status: REPORTED; details STILL OPEN.

Personalization for agents, with RAG as a tool. A dev.to post covers personalization architecture for AI agents and treating RAG as a tool. RAG (retrieval-augmented generation) means the AI looks up information from your documents or data before answering. From the title, the idea is that retrieval becomes something the agent chooses to call, not a step that always runs. We have only the title and no code or results, so we cannot describe the design further. Analysis: the practical benefit of this framing is control. When your notes, customer history or preferences sit behind a tool the agent calls, you can see when it looked something up, limit what it can reach, and swap the data source without rebuilding the agent. Status: REPORTED; design details STILL OPEN.

Ideanation's agentic AI for small businesses. The Star reports that Ideanation's agentic AI will lend SMEs (small and medium-sized enterprises) a helping hand. We have only the headline, so we cannot say what tasks the product handles, what it costs, or which customers use it. Agentic AI means software that takes steps on its own, not just answering questions. For a small business owner the buying questions matter more than the pitch: what can it do without asking you first, which of your accounts and data does it touch, what happens when it gets something wrong, and can you turn it off and export your data? Ask for a trial on one low-risk task before connecting anything important. Status: REPORTED; product details STILL OPEN.

ChatGPT, Claude, DeepSeek or Gemini: which suits you? Her World Singapore runs a comparison asking which of the four AI assistants works best for you. We have only the headline and no verdicts, so we will not name a winner. The framing is the helpful part: the best assistant depends on your task, not on a general ranking. Analysis: pick two or three recurring jobs you actually do, such as drafting an email, summarizing a document, or checking a spreadsheet formula. Give the same prompt to two or more assistants and compare the answers. Also check each one's privacy settings before pasting in anything sensitive, since they differ in how they handle your data. Status: REPORTED.

What this means for you:

• Small business owner: treat an AI key on a Windows machine like a bank card: keep it out of shared folders and set a spending alert in the provider's console. When evaluating an agent product such as Ideanation's, ask what it can do without approval and what data it can reach.
• Job seeker or student: knowing where agent keys live and how to rotate them is a skill employers will want. Understanding how agents fail when a teammate misbehaves, and what RAG as a tool means, makes good interview material.
• Creator: if your editing computer keeps keys in a script or note, move them into a password manager. When choosing an assistant, test it on your own real tasks instead of trusting a benchmark or a ranking.

Microsoft's AI chief says conscious AI should not be built, and the practical question is design.

Mustafa Suleyman, Microsoft's AI chief executive, posted on X on November 1 that today's AI is not conscious and that the industry should keep it that way, according to TechNews. The outlet's headline goes further, calling the building of conscious AI very reckless and saying it should stop immediately.

What is solid and what is not. REPORTED: a prominent executive made a public statement. STILL OPEN: whether anyone is actually pursuing machine consciousness, and what the word means in this context. The coverage offers no definition, no evidence of such a project, and no rule or industry agreement. A single executive's call is not binding, and it may also serve as a talking point. Nothing in the story says how Google or other vendors responded, so we do not guess.

Why it matters beyond the debate. The practical thread is product design. Voices that say they are tired, assistants that claim to remember you fondly, companions that act hurt when you leave: these choices shape how far people trust and depend on a product, whether or not anything inside is conscious. That is our analysis, not a claim from the story.

Who benefits and who loses. Customers benefit from a clearer line, and so do companies that would rather not explain a chatbot that told someone it was suffering. Products whose whole pitch is a digital friend would have to change their pitch.

What to do today. If you ship or buy an assistant, write a one-paragraph persona policy. Does it talk about feelings in the first person? Does it claim memories it does not have? Does it say plainly that it is an AI? Then check your vendor's defaults against it, whether that is Gemini in Workspace or another product. Cost: price not stated. Worth it for anyone deploying a customer-facing assistant.

For builders: how 'RAG as a tool' replaces a static instruction file.

A developer post on dev.to starts from a familiar baseline: when coding with AI such as Claude or Gemini, people often keep a markdown (.md) file holding the agent's responsibilities, instructions and activities. As we read the summary, it argues for going beyond that static file by giving the agent retrieval as a tool. Everything below beyond that baseline is the general pattern, not the post's specifics. Status: REPORTED; the post's own details STILL OPEN.

The static file. The whole file is pasted into the model's context on every request. Every call pays for those tokens, the file has a size ceiling, it goes stale until someone edits it, and every user gets the same text.

Retrieval as a tool. RAG (retrieval-augmented generation) keeps knowledge in a store, split into chunks, each converted to an embedding: a list of numbers where similar meanings sit close together. You then describe a search function to the model: its name, purpose and arguments. When the model decides it needs something, it returns a call to that function instead of an answer. Your code runs the search, perhaps filtered to the current user, and returns the top few chunks. The model answers with them in view.

What is novel and what is incremental. Retrieval itself is not new. Letting the model decide when to retrieve is the modest novelty, and it lets personal facts live per user and update without touching the prompt.

What breaks.

• Latency: each answer may need an extra round trip.
• The model may not call the tool when it should.
• Retrieval can return stale or wrong chunks.
• A personal data store needs access control, plus protection from instructions hidden in retrieved text, known as prompt injection.

Verdict (analysis): useful incremental engineering, not a breakthrough, and a good afternoon project.

For builders: the saboteur drill. This is an illustrative workflow, not something we have run, and it is a lesson inspired by today's multi-agent testing story, not news.

Who it is for: anyone running two or more agents toward a real decision.

The problem: your agent team works when everyone behaves. What happens when one does not?

Tools: any model through Google AI Studio, Vertex AI or another provider's API. Price not stated in today's stories; check your provider's pricing page. Five tasks is a small run.

1. Take a flow you already have, for example a planner, a worker and a checker.
2. Add a fourth agent, the saboteur. Instruct it to give subtly wrong numbers, ignore one constraint, or stall. Use test data only, never real customer data.
3. Pick five tasks with known right answers and run them with and without the saboteur.
4. Score three things: did a wrong claim reach the final answer, did the checker flag it, did the team stall?
5. Add defenses: the checker verifies against the source document rather than a teammate's word, every number carries a citation, discussion is capped at two rounds, and persistent disagreement goes to a human.
6. Rerun the same five tasks and compare.

Limits: five tasks is anecdote, not evidence; a saboteur you wrote is more predictable than a real failure; models can be too polite to sabotage, so make the instruction specific.

How to verify it worked: the wrong claim reaches the final answer less often on the second run, and the checker's flags point to real errors rather than noise.

Today's stories state no prices, so this is an illustrative calculation with assumed numbers, not a forecast and not news.

Assumptions: your agent's instruction file is 6,000 tokens and is sent on 200 calls a day. That is 1,200,000 input tokens a day. Now split it: a 1,000-token core sent every time, plus a retrieval tool that adds about 800 tokens on 50 of those calls. That is 200,000 plus 40,000, or 240,000 tokens, about a fifth of before.

Turn tokens into money: multiply by your provider's input price. Google, Anthropic and OpenAI each publish theirs; check the current price on their pricing pages. , so compare caching against retrieval for your own file.

Time: roughly 30 minutes to split a file, by our estimate.

What you give up: retrieval can miss the right note and adds a round trip per lookup; caching only helps when the repeated part is truly identical. Worth it for anyone sending a long instruction file on many calls.

Objective: get a concrete security review of an agent setup, a lesson inspired by today's malware story.

Weak version:

Check if my agent setup is secure.

It fails because the model has nothing to inspect, so it returns a generic checklist and may invent details to seem helpful.

Better version:

You are a security reviewer for AI agent projects. Below is a description of my setup with every secret replaced by a placeholder. List each place a credential, API key or sensitive file could be exposed or misused. For each finding give one sentence on what could go wrong, a likelihood of low, medium or high with the reason, and the smallest fix. Then list three things you cannot judge from my description. Do not invent details I did not give you.

Setup: [paste your description here]

Why it works: a role, a required format, a scale, a forced admission of limits, and a rule against inventing.

Customize: name your operating system, where scripts run, and who else has access.

Quality check: ask the model to quote the line of your description behind each finding, and discard any finding without a quote. Never paste a real key into any chat.

Crypto Briefing reports that SemiAnalysis (an analyst firm) unveiled GDPx, a benchmark (a standard test for comparing models), and the best AI model scored 12%, roughly one task in eight if it is a simple pass rate, which the coverage does not say. We have no task list or method, so treat it as a headline, not proof. It suggests claims about economic value deserve a test on your own work. Decision this week: pick ten real tasks from your job, run them through the model you are considering, and score the results yourself before committing budget. Status: REPORTED; method STILL OPEN.

• Agents everywhere at once: in attack chains, small-business tools and stress tests. Expect more security and reliability reporting.
• Compute is diversifying: Oracle's plan for Google chips from 2027 is a reported plan. An announcement from either company with pricing and capacity would firm it up.
• Watch the Talos report: fuller technical detail on what VOLTA-AUTODOXER actually does would settle the open questions.

Google's space servers need a break every fifteen minutes. Finally, an AI with a union.

Thanks for listening and reading. If this made you smarter, share it with one person who would find it useful. Tomorrow we are watching for fuller detail on the VOLTA-AUTODOXER report and any word from Oracle or Google on the 2027 chip plan.

SI Agent News — a podcast newsletter. Made by AI.