AI Product Updates, August 5: Hy3 Goes Global, Wispr Flow Adds a Notetaker, and Agent Controls Tighten (2026)

Coverage window

This issue covers AI product announcements and state changes published on August 5, 2026, 00:00-23:59 (UTC+8). The confirmed set is lighter on new frontier models and heavier on access, workflow capture, and control over agents.

At a glance

Product or platform	Releasing entity	What changed	Availability	What to check
Tencent Hy3	Tencent	The July 6 model release received a global-access expansion, with open weights, developer entry points, and a published OpenRouter rate card. 1	WorkBuddy is free worldwide until August 31 Pacific Time; API, TokenHub, Miora, Hugging Face, ModelScope, and third-party routes are listed.	Test tool use, latency, and license fit before routing production traffic.
Wispr Flow Notetaker	Wispr Flow	Mac users can capture system audio without adding a meeting bot, then get a live transcript, summaries, action items, and cross-meeting queries. 2	Available on Mac; included with Wispr Flow. Windows is coming, mobile is on the roadmap, and English is the current language.	Review participant notice, retention, and MCP permissions before enabling it in real meetings.
Aegisora	Aegisora	A new open-source proxy says it can intercept agent actions, enforce least-privilege API access, mask PII, and produce audit logs. 3	Product Hunt launch page says "Launching today" and lists free options; a separate paid plan is not disclosed.	Verify the claimed zero-latency path and test policy failures, not just successful calls.
Android AI interoperability	Google Android	Google published a position on EU DMA measures that could give downloaded AI agents deep access to device capabilities. This is a policy update, not a shipped Android feature. 4	Applies to the European regulatory discussion; no new public endpoint or general-availability switch was announced.	Treat microphone, camera, screen content, and screen automation as high-risk permissions.

Model and access

Tencent Hy3 is now easier to reach globally

Tencent's August 5 post is an access expansion, not a new model version. Hy3, originally released on July 6, is described as a 295-billion-parameter Mixture-of-Experts model with 21 billion active parameters and a context window of up to 256K tokens. Tencent says the weights use the Apache 2.0 license. 1

The practical change is the number of doors. Tencent lists WorkBuddy, Tencent Design Miora, Tencent Cloud TokenHub, direct API access, third-party tools, Hugging Face, and ModelScope. WorkBuddy is free worldwide through August 31, 2026, Pacific Time. The post also identifies regional partners and developer integrations, so the model is no longer confined to a single Tencent surface.

The same announcement gives an OpenRouter rate card of $0.1288 per million input tokens and $0.5336 per million output tokens. It does not state the previous rate, so this should be read as a current listed price, not as a verified price cut or increase. 1

Tencent reports more than 68 times the API-call volume of the previous Hy generation, first place on OpenRouter's global usage leaderboard within one week, more than 90% task success in internal WorkBuddy evaluations, and a 34% reduction in average task time. These are company-reported figures, not independent benchmarks. The useful question for a developer is narrower: does Hy3's agent and coding behavior survive your tool schema, context size, and latency budget?

Action: run the same tool-calling and long-context test suite against Hy3 and the model already serving your workflow. Record time to first token, total completion time, tool-call error rate, and output quality. The Apache license broadens deployment options, but it does not remove the need to validate the serving stack and regional availability.

Workflow products

Wispr Flow turns meeting audio into a searchable work record

Wispr Flow officially launched its Notetaker for Mac on August 5, according to TechCrunch's update to its report. It uses system audio, so it can transcribe a meeting without joining the call as a visible bot. The product exposes a live transcript, cleans up the text, creates action items, and lets users query across meeting history. 2

Wispr's own launch page adds the workflow detail: speaker attribution, a "What did I miss?" catch-up mode, summaries organized around decisions, next steps, dependencies, blockers, and open questions, and connections to Claude, ChatGPT, Cursor, and other MCP-compatible tools. It works with Google Meet, Teams, Huddles, and unplanned or in-person conversations captured by the Mac microphone. 5

Availability is Mac today, with Windows coming soon, mobile on the roadmap, and English support first. Notetaker is included with Wispr Flow; the official page says the free plan remains free forever and Pro includes unlimited use. Wispr says audio is encrypted, retained only temporarily, automatically deleted after the limited retention period, and not used to train models without consent. It also says users can delete meetings, transcripts, and notes. 5

The operational caveat is easy to miss: the product may capture people who did not install it. Wispr says users are responsible for informing participants before recording. That makes consent and retention part of the product setup, not an administrative afterthought. MCP access also changes the risk boundary: a meeting transcript that can be searched by another assistant is more useful, but it is no longer confined to the notetaker.

Action: before enabling it for customer or employee calls, define the participant notice, the retention period, the deletion path, and which MCP tools can read meeting history. Test an unplanned call and a speaker correction, then verify that the resulting notes do not leak into a broader assistant context by default.

Agent control

Aegisora puts a proxy between agents and tools

Aegisora launched on Product Hunt with a deliberately narrow pitch: it is an open-source proxy layer for AppSec teams that intercepts LLM actions, enforces least-privilege API access, masks personally identifiable information, and generates readable audit logs for autonomous agents. The launch page labels the product "Launching today" and lists free options. 3

The important distinction is between a control point and an agent framework. Aegisora does not present a new model or a new agent runtime. It sits in the request path and tries to make tool use inspectable and constrainable. That is the part many agent deployments still lack: a place to deny a dangerous call before the model's plan becomes an external side effect.

The page calls the proxy "zero-latency," but no independent latency measurement is provided. Treat that phrase as a product claim. For a production evaluation, measure added latency, policy evaluation time, failure behavior when the proxy is unavailable, and whether audit records contain enough context to reconstruct a blocked action without storing more sensitive data than the original request.

Action: put Aegisora in front of a non-production agent and replay three cases: a prompt-injection attempt that asks for a new tool, a request containing an email address or token, and a legitimate multi-step workflow. Check whether the proxy blocks the first two, preserves the third, and leaves an audit trail that a reviewer can actually use.

Google argues for stronger Android boundaries around AI agents

Google's August 5 Android security post is not a feature launch. It is a public argument about how the European Union's Digital Markets Act should apply to AI services on Android. Google says the European Commission's measures could give user-downloaded AI agents deep access to device capabilities, including the microphone, camera, on-screen content, and screen automation. 4

Google says it supports interoperability but wants a qualification process and platform-controlled certification lifecycle for agents, including review, approval, suspension, and revocation. It argues that open-ended access would enlarge the attack surface for prompt injection, scams, malware, and privacy abuse, and says the DMA measures may collide with other EU cybersecurity frameworks.

For product teams, the immediate status is policy discussion, not an Android API change. There is no new permission grant to adopt from this post. The signal is about the eventual contract between an agent, the operating system, and the user: who can approve a high-risk capability, who can revoke it, and whether a third-party agent can inherit the same trust as a platform-integrated one.

Action: treat Android agent integrations as capability-specific rather than blanket-authorized. Request microphone, camera, screen-reading, and screen-control access separately, log the user's grant, and make revocation visible. That is a prudent engineering response to the risks Google describes; it is not a claim that the EU has settled the final implementation.

What to test today

Model access: compare Hy3's tool calling and 256K-context behavior with the model already in production. Keep Tencent's internal success and speed figures separate from your own measurements.
Meeting capture: validate participant consent, audio retention, deletion, speaker correction, and MCP scope in Wispr Flow before using it for sensitive calls.
Agent safety: measure Aegisora's added latency and fail-closed behavior under prompt injection, PII, and proxy-outage tests. Its "zero-latency" wording is not an independent benchmark.
Mobile permissions: for Android agents in EU-facing products, design explicit, revocable permissions for microphone, camera, screen content, and screen automation while the regulatory details remain unsettled.

The common thread is operational: August 5 produced no confirmed new OpenAI, Anthropic, or Gemini frontier-model version in the primary pages checked for this issue. The useful changes were the paths around models, meetings, and tools: wider access, more retained context, and more pressure to put a human-readable control layer between an agent and the world.