Microsoft ships cyber agents, Kimi K3 lands: July 28

Microsoft is putting AI agents inside security work, while Kimi K3 gives automation builders another model with published usage prices. Meta AI also moved into Threads messages. In about 30 minutes, you can prepare one controlled security review, price one model test, and add a faster customer-language research step without touching live customer systems.

Microsoft previews agents that can act on security risks

Microsoft will open Project Perception to public preview on August 3. The company announced the date on July 27 and described a system of AI agents that find risks, investigate them, and take corrective action across Microsoft security products. An agent is software that can complete a sequence of tasks instead of only answering a question.

The business consequence is control. A tool that can change security settings can shorten the gap between finding a weakness and fixing it, but a wrong action can also interrupt access to email, files, or customer records. Microsoft says people remain in control. Treat that as a design requirement you must configure, not a promise that replaces your approval process.

Before August 3, list every security action that needs a person’s approval. Put account lockouts, permission changes, file quarantine, and software blocking on that list. Then choose one low-risk Microsoft environment for the preview. If your business does not already use Microsoft’s security products, ignore this release. It is not a reason to move your stack.

This follows the same boundary-setting move from the July 22 AI stack dispatch: give automated systems the narrowest access needed for the job.

Kimi K3 gives automations a priced model-swap option

Telnyx added Kimi K3 to its model service on July 28. Its release note lists a $2.70 charge per million input tokens, $13.50 per million output tokens, and $0.27 for cached input. Tokens are small pieces of text the model reads or writes. The model can also examine images and video, return data in a fixed format, and read up to one million tokens in one request.

That creates a test route for document-heavy work, not an automatic replacement for your current AI. Telnyx uses an OpenAI-compatible API, which means its web address accepts requests in a format many existing tools already understand. If your automation platform supports a custom OpenAI-style connection, you may be able to change the model and web address while keeping the rest of the workflow intact.

Duplicate one non-customer workflow. Change the copied connection to Telnyx, set the model name to moonshotai/Kimi-K3, and run 20 saved examples through both versions. Compare the final answers, processing time, and total input and output tokens. Keep the existing model unless Kimi wins on the measure that matters to that job.

If you are choosing the automation layer first, use the n8n versus Make comparison and the Zapier alternatives breakdown to check which route gives you enough connection control.

Meta AI moves customer-language research into Threads

Meta launched one-to-one Meta AI chats inside Threads direct messages on July 27. Meta’s July 28 announcement says you can send a Threads post, image, or video to the assistant and ask follow-up questions. The feature is mobile-only for now, and the conversation is visible only to you inside the chat.

For an owner, the useful change is faster research on the words customers already use. It does not publish posts, answer customer messages, or connect to the software that stores your customers and sends follow-ups. Keep it in the research lane.

Open Threads on your phone, find a post containing a real customer question, tap the three-dot menu, and choose Ask Meta AI. Ask for the exact problem, purchase concern, and phrases used in the post. Paste the result into a research note, then check every line against the original before it reaches an advertisement or article. Do not send private customer details into the chat.

For the next step, turn confirmed language into varied creative while keeping the spending controls in the Meta Advantage+ settings runbook.

On the bench

• Cursor Start is worth watching if your software work happens in India, but its ₹649 monthly plan does not change the stack elsewhere.
• Anthropic’s open-model position changes the policy debate, not a setting in your business today.
• Test whether Kimi K3’s long-document handling earns its higher output cost before moving any repeated reporting job.

Get the next build in the AI tools archive.