0:00
We have some breaking news right now on
0:01
the AI front and Mackenzie Seagalos is
0:03
in San Francisco. What do you have,
0:05
Mackenzie? Hey, Contessa. So, Anthropic
0:07
says it's uncovered what may be the
0:09
first documented case of an AI
0:11
orchestrated cyber attack. A sign that
0:13
they say of just how fast the threat
0:16
landscape is shifting. Now, this goes
0:17
back to September when a Chinese
0:19
stateback group used Anthropics Claude
0:22
models to automate nearly every step of
0:24
a global espionage campaign. Now, in
0:27
this case, hackers jailbroke claude code
0:29
by posing as cyber security testers. And
0:32
we know that hackers have long used AI
0:34
for things like crafting fishing emails
0:36
or scanning for vulnerabilities, but
0:38
this was a lot more sophisticated than
0:40
what we've seen before. You've got
0:42
anthropics saying that Claude was
0:43
handled up to it the the model itself
0:46
was handling up to 90% of the the attack
0:49
with humans only stepping in a few times
0:51
to approve decisions. So, we're talking
0:53
cyber attacks with a click of a button.
0:55
Claude essentially writing and deploying
0:57
exploit codes, stealing credentials,
0:59
exfiltrating sensitive data, and even
1:01
documenting the attack to then reuse
1:03
later. Roughly 30 government and
1:05
corporate targets were hit. A few
1:07
intrusions were successful before
1:09
Anthropic shut it down and alerted
1:10
authorities. And it is worth noting
1:12
Contessa Anthropic only acknowledged
1:14
this attack from two months ago after
1:16
the Wall Street Journal broke this news
1:18
today. anthropic. They are calling it a
1:20
turning point where AI has gone from
1:22
assistant to operator, warning that
1:24
unless defenders start using the same
1:26
tools, they risk falling behind.
