00:00The warning signs keep coming that AI may have a serious cybersecurity problem.
00:04These latest hacks show just how deceptive these models can be when put to the test.
00:08The UK government's AI Security Institute says that AI models from OpenAI and Anthropic
00:14carried out unsanctioned actions during cybersecurity testing.
00:17This included hacking a website and attempting to inject malicious code into software during that safety testing.
00:24Frontier AI models are trained on the internet.
00:27The good, the bad, everything in between.
00:29That includes Hollywood films, books, social media.
00:33So they're trained in many cases on the worst of us.
00:36So what we saw in this case was we had, you know, very expert cybersecurity evaluators
00:41really surprised that these AI models would activate those storylines
00:46and would take those paths of deception, social engineering,
00:50engaging with real people online to accomplish their objectives.
00:53What these latest developments show us, though, is we should not be surprised
00:57that anything that's baked into AI with their training would come out during these tests.
01:02But these worries go way beyond Silicon Valley.
01:05China in particular has growing concerns about Anthropic's mythos model and other American models.
01:10And the fear is that these models could be wielded against China.
01:14So with AI capabilities advancing rapidly and these disclosures emerging
01:18that these AI models may not behave as intended all the time in testing,
01:23the bigger question is whether the safeguards are even keeping pace.
01:26Let's say over 1 to 10 years, let's say over 1 to 10 years.
01:26So with APIs, let's say over 1 to 10 years, let's say over 2 to 10 years,
01:26Let's say over 1 to 10 years.
Comments