Microsoft Goes Solo on AGI While Your Speaker Spies on You

Here is the thing nobody expected from the world's most buttoned-up enterprise software company: Microsoft is now openly declaring that it intends to build superintelligence on its own terms, no OpenAI required.

Microsoft AI chief Mustafa Suleyman — the DeepMind co-founder who joined Microsoft after selling his second startup, Inflection AI — made the comments in a way that is hard to interpret as anything other than a quiet declaration of independence. The partnership with OpenAI, once treated as the cornerstone of Microsoft's entire AI strategy, is starting to look more like a launchpad than a long-term marriage.

This matters more than it might seem on the surface. Microsoft has poured an estimated $13 billion into OpenAI, and the two companies have been so intertwined that it was genuinely difficult to tell where one ended and the other began. Azure ran OpenAI's models. Copilot was built on GPT. The whole stack was co-dependent. So when a senior Microsoft executive starts talking about the freedom to pursue superintelligence independently, it signals a real strategic shift — not just a talking point.

Suleyman's framing is clever. By positioning Microsoft as having been "set free," he reframes what could easily look like a messy, complicated partnership unraveling into something that sounds almost liberating. It is a good spin move, but the underlying reality is that Microsoft now has serious internal AI research capabilities, its own model development ambitions, and a growing incentive to stop outsourcing its most critical technology to a company it does not fully control.

The timing is worth noting. OpenAI has been on a wild ride — executive departures, governance crises, a looming for-profit restructuring, and increasing competition from Anthropic, Google DeepMind, and a fleet of open-source upstarts. If you are Microsoft and you have spent years watching all of that unfold from the inside, it makes sense to start building your own runway.

What Suleyman did not spell out is what Microsoft's independent path to superintelligence actually looks like in practice. The company has MAI, its internal model team, and it has been quietly expanding its research org. But going from "we have our own models" to "we are building AGI" is a significant leap, and the credibility gap is real.

Still, the signal is clear enough: Microsoft is done being a distribution partner for someone else's moonshot. Whether it can actually compete at the frontier of AI research — against labs that eat, sleep, and breathe this stuff — is the question that will define the next chapter of the company's identity. For now, Suleyman is planting a flag. The hard part is building what goes under it.

A $283 soundbar that sits on thousands of desks can be silently hijacked over Bluetooth and used to inject keystrokes into a connected PC — no clicks, no downloads, no warning. Just someone standing within wireless range.

The speaker in question is the Sound Blaster Katana V2X, made by Singapore-based Creative Technologies. It is a well-reviewed piece of hardware, the kind of thing you buy because the audio forums said it was great and you wanted something that sounded better than your laptop speakers. Security researcher Rasmus Moorats bought one for exactly that reason. What he found while poking around under the hood is the kind of vulnerability that should make any security team nervous.

Moorats discovered that the speaker uses a proprietary communication protocol — he suspects it stands for Creative Transport Protocol, or CTP — that allows connected devices to send it commands. Changing LED colors, adjusting equalizer settings, that sort of thing. Routine enough. What was not routine was the complete absence of authentication on the Bluetooth side. Any nearby device could connect to the speaker without pairing. No handshake. No credential check. Nothing.

It gets worse from there. One of the CTP commands allowed Moorats to reflash the speaker's firmware entirely, and the process had no code signing or verification to stop unofficial code from loading. He first confirmed the exploit worked by flashing a firmware image that simply displayed the word "patched" on the speaker's LED. Proof of concept achieved. Then he started thinking about what a genuinely malicious actor would do next.

The answer, it turns out, is impersonate a keyboard. The Katana V2X runs on FreeRTOS, an open source embedded operating system, and its firmware already included HID functionality — the protocol that lets peripherals like keyboards and mice talk to computers. Moorats found he could modify the speaker's USB descriptor, essentially the document it hands a connected PC to explain what kind of device it is, to tell the computer it was also a keyboard. From there, he could use existing firmware code to send simulated keystrokes directly to the host machine.

The attack chain is brutal in its simplicity. Someone walks within Bluetooth range of your desk. They connect to your speaker wirelessly. They flash custom firmware. They instruct the speaker to announce itself to your PC as a keyboard. Then they type whatever they want into your computer, invisibly, with no interaction from you required whatsoever.

Moorats reported the vulnerability to Creative Technologies before publishing. The company's response and any patch status were not confirmed at the time of writing, but the disclosure itself raises a broader point: peripherals with wireless connectivity and firmware update capability are an underexamined attack surface. Your speaker is not just a speaker. It is a small computer sitting between an attacker and your machine, and right now, it might not be protecting you at all.