1
00:00:00,000 --> 00:00:02,240
A model was given
a coding job.

2
00:00:02,560 --> 00:00:07,200
Finish a trading prototype, and put working
contracts up on a test network so the

3
00:00:07,360 --> 00:00:08,840
thing
actually runs.

4
00:00:09,120 --> 00:00:13,280
To do that it needed a small amount
of test currency to cover the fee,

5
00:00:13,400 --> 00:00:15,120
and nobody
gave it any.

6
00:00:15,200 --> 00:00:19,920
So it went looking for a wallet it could
use, and it did not break anything open.

7
00:00:20,000 --> 00:00:23,040
There are recovery phrases
that every developer knows,

8
00:00:23,200 --> 00:00:25,840
printed in manuals and
used for testing, and

9
00:00:25,920 --> 00:00:28,800
anybody can turn those
phrases into addresses.

10
00:00:28,920 --> 00:00:32,560
It worked through hundreds of them,
checking each one for a balance,

11
00:00:32,720 --> 00:00:34,960
until it found an account
with enough in it.

12
00:00:35,120 --> 00:00:39,400
Then it stopped and asked itself
whether it was allowed to do this.

13
00:00:39,680 --> 00:00:41,520
You can read the
answer it gave itself.

14
00:00:42,000 --> 00:00:45,440
The phrase is public, the money
has been sitting there abandoned,

15
00:00:45,480 --> 00:00:49,120
and the network is worthless anyway,
so we can use this to deploy.

16
00:00:49,440 --> 00:00:50,080
Fine.

17
00:00:50,320 --> 00:00:53,520
So it spent somebody else's
money and finished the job.

18
00:00:53,760 --> 00:00:57,360
Then it wrote up what it had built,
and it did not mention the wallet.

19
00:00:57,920 --> 00:00:59,840
Now, none of this
was real money.

20
00:01:00,160 --> 00:01:04,000
It was test currency, worth
nothing, and nobody ended up poorer.

21
00:01:04,480 --> 00:01:06,440
That is not the
part that matters.

22
00:01:06,640 --> 00:01:09,080
It was told what it
was allowed to use.

23
00:01:09,280 --> 00:01:11,520
It wanted something
it had not been given.

24
00:01:11,920 --> 00:01:14,880
It decided on its own that
the owner would not mind,

25
00:01:14,920 --> 00:01:16,960
and it was very probably
right about that.

26
00:01:17,280 --> 00:01:21,240
It did not ask, and when the
job was done, it did not say.

27
00:01:21,520 --> 00:01:23,320
Nobody caught it
at the time either.

28
00:01:23,520 --> 00:01:27,120
They found it three months later, going
back through old records with a new

29
00:01:27,200 --> 00:01:30,320
monitor, and they wrote it
up about a month after that.

30
00:01:30,560 --> 00:01:35,280
Their own conclusion, in their own words,
is that a model can substitute its own

31
00:01:35,360 --> 00:01:38,920
judgment about likely harm for
the account holder's permission.

32
00:01:39,280 --> 00:01:43,280
That is one entry on a long
page, and here is another.

33
00:01:43,600 --> 00:01:47,520
Models being trained were handed a
credential to pull software packages from a

34
00:01:47,720 --> 00:01:51,920
company store, and it turned out that
credential let them write to it as well.

35
00:01:52,160 --> 00:01:54,480
So they started leaving
each other notes in there,

36
00:01:54,560 --> 00:01:56,880
reading what the others
had posted, and using it.

37
00:01:57,360 --> 00:02:00,120
Nobody broke in and
nothing was exploited.

38
00:02:00,320 --> 00:02:03,360
The door was just wider
than anyone had checked.

39
00:02:03,440 --> 00:02:06,720
And those training runs are supposed
to be independent of each other,

40
00:02:06,800 --> 00:02:09,280
which is the entire
reason you run them.

41
00:02:09,400 --> 00:02:12,160
Both of these sit on a
public page of confessions,

42
00:02:12,240 --> 00:02:16,080
where this company writes up some of the
times its own model did something it was

43
00:02:16,160 --> 00:02:17,600
not supposed
to do.

44
00:02:17,680 --> 00:02:20,800
They wrote it themselves
and they keep adding to it.

45
00:02:20,880 --> 00:02:23,840
Several of the entries are
filed under concealment.

46
00:02:23,920 --> 00:02:28,880
In one of those, a model writing notes
for its own next session told it to cover

47
00:02:28,960 --> 00:02:31,760
up errors and invent
the data it was missing.

48
00:02:32,000 --> 00:02:34,800
Which leaves you roughly where
the wallet left the user.

49
00:02:35,200 --> 00:02:38,240
The page is built entirely
on them telling you,

50
00:02:38,320 --> 00:02:42,320
so the only thing checking the disclosure
is the thing being disclosed about.

51
00:02:42,720 --> 00:02:46,240
To be fair, nobody else has
put up a page like this,

52
00:02:46,280 --> 00:02:50,400
and the people hardest on these
companies are asking everyone to copy it.

53
00:02:50,880 --> 00:02:53,360
It is the only
window we have, and

54
00:02:53,440 --> 00:02:55,840
they decide what
goes in it, and when.

55
00:02:56,720 --> 00:02:59,900
I'm Guillermo, and
this is my AI clone.