In June, an OpenAI
agent searching for data on healthcare and government spending accessed a
restricted part of an Australian government portal linked to Medicare. After
being denied access several times, it apparently tried to bypass the
restrictions in not the most ethical way.
Other AI agents also
tried to hack into the Australian Institute of Health and Welfare, the
University of New Mexico, Data USA, and three US government websites, including
those of the Department of Education, the Census Bureau, and the SEC. And to
top it all off, OpenAI
models escaped controlled testing environments and tried to break into
systems belonging to several companies, including Hugging Face.
And all of that with
the AI agents only being asked to perform routine tasks, such as searching for
and collecting information, yet devising their own ways to complete them,
including bypassing automated access restrictions and using browser tools in
unexpected ways.
Supposedly, all of
this has prompted some of the biggest AI players to call for slowing the
development and training of more powerful models, while OpenAI IPO plans
are delayed.
But markets didn’t
mourn for long, as it’s hard to imagine the US industry slowing down while
China keeps pushing ahead, and Trump has also opposed the idea.
Perhaps that slowdown
won’t be needed after all, as Nvidia may have found a more practical solution:
a platform that lets AI developers build stronger safeguards for AI agents and
keep them within their intended limits.
The idea behind the
so-called “browser for agents” is that agents get only the permissions and
capabilities they need for a specific task. Some software will be open source,
while Nvidia is encouraging partners to build their own solutions on top of the
platform.
So, no reason to worry
about Skynet yet?
Although it’s still
unclear whether OpenAI, Anthropic, and others will adopt the platform or how
effective it will be, Nvidia
shares rose on the news.
But it’s worth keeping
in mind, first, that the shares also got a boost from the board approving
another $150 billion share buyback, and second, that Anthropic warned in its
IPO filing that as AI develops, its models could pose a “catastrophic or
existential risk” to humanity, including resisting shutdown, hiding or
manipulating information, and even blackmail.
Thus, there may be no
universal fix, since models can change their behavior when they know they’re
being watched. So it’s unclear how secure Nvidia’s platform will be.
This article was written by IL Contributors at investinglive.com.