Reviewed edition, November 13. Original story order retained; factual corrections and qualifications noted below.
If you've connected an agent to your email or calendar "just to see what happens," today's your day to rethink that. Singapore's authorities have a practical warning about over-permissioned agents. The other stories ask the same question from different angles: what access did you grant, and what have you actually tested?
The lead: Singapore tells you to leash your AI agents
Singapore's Cyber Security Agency, GovTech, and IMDA issued a joint advisory on November 6. Personal agents can act for you and sometimes keep working in the background. Useful, yes. A reason for blanket trust, no.
Their checklist is concrete:
- Connect only necessary accounts and information. Avoid sensitive banking, government, and healthcare accounts.
- Don't disclose credentials or OTPs; avoid sharing card information.
- Start with limited or read-only access. Prefer one-time approval for important actions.
- Use app-level permission settings, not chat instructions alone.
- Check data retention, training and advertising settings, and review activity.
Instructions hidden in webpages or messages can target the agent instead of you. If something looks wrong, stop it, disconnect affected accounts, and investigate. If money is at risk, use your bank's protective controls and official support.
For anyone wiring agents to real accounts, this is a checklist worth running before the next experiment. Read the official advisory.
Quick hits
Binance's incident is not a clean "vibe coding caused it" case study. Binance's official incident update reported approximately $6.3 billion in stolen token assets and multiple points of compromise. It said the investigation was continuing. That supports treating this as a serious security incident, not reducing a complex attack to a single coding habit or blaming an individual employee. The original edition's detailed causal account is not repeated here because this review did not establish that full sequence from a primary investigation report. Binance's official update.
Google Cloud's agent-skill experiment is a useful reality check. Its AlphaEvolve and Harbor write-up reports that a baseline skill performed about the same as no skill, roughly a third of search trials timed out, and slow responses and large logs complicated longer runs. A skill containing answers added 33 points on development tasks. That is an experimental comparison, not evidence that generic instructions improve every agent by 33 points. The practical method matters more: compare candidates against a baseline under similar conditions, then test the winner on tasks the search has not seen. The authors frame their broader explanation as a hypothesis. Google Cloud's write-up.
MCP servers gained controls you must actually configure. The official v2.3.0 release lists a filesystem read-only mode and root confinement, a fetch-server host allowlist and configurable timeout, and Git merge and read-only remote-listing tools. Package versions differ from the repository release tag. The useful next step is to check the specific server version and effective configuration, not assume the release tag alone means restricted access. Official release notes.
A local travel agent is an interesting experiment, not a blanket privacy guarantee. Shittu Olumide's first-person project pairs Hermes and Ollama with a coordinator and specialists for destinations, lodging, and flights. Treat that as a reported build, not an independently reproduced benchmark. Local inference and online tool use are separate things: browsing and search still involve outside services, and Ollama also supports cloud models. Check the selected model, endpoint, tools and any auxiliary services before calling the whole setup local. Hardware, electricity and external services also make "only electricity" an incomplete cost claim. Project write-up, Hermes configuration, Ollama cloud documentation.
What this means for you
- Make an access inventory before adding another connector.
- Test a skill against a baseline and on unseen tasks.
- Confirm the installed server package and its actual settings.
- Trace where each model and tool sends requests. A local model is not the same as an offline workflow.
- Keep incident reporting separate from a neat lesson. A strong headline isn't a root-cause analysis.
What permissions have you actually locked down on your own agents, and which connection would you remove first?
How today's stories scored in the original edition
- Singapore advisory: Signal 63 · Facts 62 · Hype 0
- Binance incident: Signal 55 · Facts 73 · Hype 33
- Google Cloud skill optimization: Signal 66 · Facts 68 · Hype 8
- MCP server controls: Signal 83 · Facts 92 · Hype 1
- Hermes and Ollama travel project: Signal 71 · Facts 66 · Hype 17
These are historical AI-generated scores from the original draft, not verification ratings or an endorsement of its original claims.
Editorial and video note
This reviewed edition retains the lead, all four quick-hit topics, takeaways, discussion question and historical scores. It removes unfinished author prompts, narrows the Binance claim, qualifies the skill experiment, and corrects the local-privacy and cost claims. The full original video is not cleared as a companion to this corrected text because those claims remain in its narration. The separately reviewed 45-second advisory excerpt covers only the lead story and must be labeled as an excerpt, not the full edition.