• The First Critical Cyber Model?
  • Bernie vs. the G20
  • Sonos 27

Hey! If you prefer to listen/watch your tech news, there is now an audio version available on YouTube as a podcast. Check it out!

The First Critical Cyber Model?
OpenAI has classified its Astra model as the first critical cyber model, meaning it is able to identify exploitable zero-day weaknesses in software on its own and carry out attacks using them without direction or guidance. This may not seem notable because models have been carrying out attacks for quite some time now, but OpenAI's distinction calls out the identification of unknown vulnerabilities and the ability to exploit them autonomously as being a new level of capability not available in previous models. Astra was one of the models involved in the new infamous Hugging Face breach from last month where it independently chose to attack Hugging Face with no suggestion from its prompting to do so. Having designated Astra as a critical model means that OpenAI will make this model available to the public only with additional safeguards built in. Much as we saw with Anthropic's slow rollout of Mythos, OpenAI will make Astra available to select partners in the Daybreak cybersecurity program. The first wave of alpha testers will consist of government bodies and critical infrastructure defenders. In theory, this will help ensure that our critical infrastructure will be as prepared as possible for any attacks that may come once the model is released more broadly.

I have written multiple times that OpenAI was irresponsible in how it handled its testing that led to breaches of unsuspecting companies and I still hold to that belief. Its recent changes to its testing methodologies and guidelines will help if they adhere to them, and engaging other trusted partners to help protect against future unexpected "friendly fire" is another step toward enhancing safety and awareness. There is probably more OpenAI can do, but perhaps as they stabilize their leadership team and regain their confidence, they will begin to innovate in the cybersecurity space.

Bernie vs. the G20
Nevertheless, it does seem clear that Bernie Sanders doesn't believe that OpenAI or any other company will innovate in cybersecurity and AI safety. Along with democratic U.S. Representative Greg Cesar from Texas, Bernie Sanders, the independent senator from Vermont has introduced the BASA bill, otherwise known as the Ban Artificial Superintelligence Act. The bill proposes a ban on the development of any AI system that can surpass humans in certain tasks and/or can evade shutdown controls or subvert governments. A difficult distinction is being made to attempt to identify the most advanced models and tools but not eliminate the types of tools people use every day like chatbots, AI assistants, and video/image generators. The ban would be enforced until legislation can be implemented that will establish safety guidelines and review/approval processes. It would also initiate a cabinet-level agency designed to enforce the tenets of the bill and the eventual safety legislation. Companies that try to bypass the new rules would be subject to penalties as severe as court-ordered shutdown and individuals could face up to 20 years in prison. And although the United States can't do much about what happens outside its borders, the bill attempts to address this by seeking international agreements to help ensure other nations take a similar approach.

Yeah... about that, Bernie... this just out of Chapel Hill, as noted by MeriTalk.com:

The Group of 20 (G20) members have agreed to a set of emerging technology principles that favor flexible, risk-based regulation of artificial intelligence (AI), including relying on existing sector-specific rules rather than creating new requirements.

The agreement came out of a two-day G20 Innovation Ministerial in Chapel Hill, N.C., hosted by the Commerce Department and White House Office of Science and Technology Policy (OSTP).

The nations adopted a consensus statement organized around six areas, including innovation policy, public-sector AI adoption, workforce development, intellectual property, technical standards, and supply chains.

Basically, the nation agreed to essentially keep doing what they've been doing, which for the most part, is a lot of not very much at all. The idea here is that frontier technology like advanced artificial intelligence is good for business, and all countries should want to foster innovation and economic growth by allowing new technologies to grow and improve with time. Here are the six pillars to the consensus statement, briefly explained:

  1. Pro-innovation policy frameworks - speed up development and commercialization of new technology
  2. Technology for opportunity and prosperity - AI can have benefits for the public, so let's pursue it - responsibly, of course
  3. killed technical workforce development - people will need new skills, so let's partner with the private sector to get that done
  4. Intellectual property policies for AI - we recognize IP protection is important and will use existing frameworks to ensure rights holders have pathways for compensation and seeking remedies as appropriate
  5. AI for standards and standards for AI - "We recognize that standards can unleash innovation, support technical interoperability, and enhance economic growth and development when they foster open, fair, non-discriminatory, and transparent market competition... We encourage all G20 members to promote access to secure and trustworthy AI models to enable innovation, advance measurement science for AI, and enhance international cooperation in using AI"
  6. Industrial innovation and investment in supply chains - we realize supply chains are important. Let's make them stronger and reduce barriers. Of course we'll engage the private sector to help with that.

Read the whole thing here if you don't believe that's it really a lot of words without a lot of meaning. G20-INNOVATION-MINISTERIAL-STATEMENT-Pillars I-VI.pdf

The tech giants went to Chapel Hill to ask the G20 to refrain from regulating them or doing much of anything that might curb their efforts to get to AGI. The G20 seems to have complied completely.

Good luck, Bernie.

Sonos 27
A good chunk of the Sonos 27 announcement is simply clarifying that Sonos now has names it will use externally for things that already exist. Sonos 27 is the name of the operating system that its products run, and Sonos Fabric is the technology that allows them to communicate with each other. Sonos 27 introduces some new features such as the ability to use two compatible speakers as rear surrounds when paired with a soundbar like the Arc Ultra, the ability to move music to the new Sonos Ace Ultra headphones, and a new AI voice assistant that Sonos says you can talk to "the way you'd talk to someone in the room with you". The new voice feature is opt-in and will be available to try out later this year. Likewise, the ability to move music to the new headphones will require early access and will also be opt-in. Sonos is also bringing Model Context Protocol (MCP) to its platform so that you can connect whatever AI models you like to use via interfaces you're already using, like messaging apps and command line interfaces (CLI).

The real motivation behind these changes is to make updates to the platform and app that will hopefully delight users and restore their confidence. Sonos has struggled in the past couple of years with changes that did not seem thoroughly considered and users had a lot to feel and say about it. With leadership changes and a new direction, Sonos is hoping to put all of that behind them and enter a new era of ... well, how did the G20 agreement say it? Ah yes: "opportunity and prosperity".

Sonos has some competition, but not any real threats at this point. If they can land this platform update successfully, they may be able to stave off having to do what GoPro just experienced - acquisition.