1
00:00:00,500 --> 00:00:01,820
Cybersecurity today.

2
00:00:01,825 --> 00:00:05,584
Would like to thank Meter for their
support in bringing you this podcast.

3
00:00:06,369 --> 00:00:11,080
Meter delivers a complete networking
stack wired, wireless and cellular

4
00:00:11,139 --> 00:00:15,639
in one integrated solution that's
built for performance and scale.

5
00:00:15,700 --> 00:00:20,080
You can find them at meter.com/cst.

6
00:00:21,640 --> 00:00:26,290
Living off the land attacks on Microsoft,
attackers use fake Calendly invites

7
00:00:26,290 --> 00:00:29,020
to hijack Google and Meta ad accounts.

8
00:00:29,290 --> 00:00:35,230
Oracle breach claims yet another victim
and researchers trace AI jailbreaks to

9
00:00:35,230 --> 00:00:37,720
hidden patterns in sentence structure.

10
00:00:38,050 --> 00:00:40,090
This is Cybersecurity Today.

11
00:00:40,450 --> 00:00:41,710
I'm your host, Jim.

12
00:00:41,710 --> 00:00:42,130
Love

13
00:00:43,900 --> 00:00:47,560
living off the land attacks is a phrase
that was first coined about a dozen

14
00:00:47,560 --> 00:00:52,870
years ago by Christopher Campbell
and Matt Grabber at Derby Con three.

15
00:00:53,140 --> 00:00:56,890
It describes a technique where attackers
can hide using existing software and

16
00:00:56,890 --> 00:01:01,153
utilities I was reading about this and
how it had been used in the Ukraine war

17
00:01:01,333 --> 00:01:06,583
where intruders were relying entirely
on Microsoft utilities to stay hidden.

18
00:01:06,673 --> 00:01:11,263
And then I looked around and I found
more and more examples of where

19
00:01:11,263 --> 00:01:15,733
intruders were relying entirely
on trusted Microsoft utilities.

20
00:01:15,973 --> 00:01:19,993
But now in North America,
The stories are the same.

21
00:01:20,113 --> 00:01:26,863
Attackers start with a foothold and they
pivot Using PowerShell WMI task scheduler

22
00:01:27,013 --> 00:01:29,473
all legitimate windows components.

23
00:01:29,833 --> 00:01:34,153
PowerShell might quietly
download a script, WMI might

24
00:01:34,153 --> 00:01:36,163
execute commands remotely.

25
00:01:36,493 --> 00:01:41,113
A scheduled task or a registry run key
might be altered to keep the attackers

26
00:01:41,113 --> 00:01:45,763
hidden inside And nothing they do
involves a suspicious binary and

27
00:01:45,763 --> 00:01:47,923
nothing looks foreign to the system.

28
00:01:48,283 --> 00:01:52,033
These are the same tools
administrators use every day, and

29
00:01:52,033 --> 00:01:56,833
they're also tools Windows relies
on, so you can't simply block them.

30
00:01:57,193 --> 00:01:58,183
And of course.

31
00:01:58,608 --> 00:02:01,728
That's why these attacks
evade so many defenses.

32
00:02:02,208 --> 00:02:05,928
Endpoint Detection and Response systems
are excellent at spotting malware,

33
00:02:06,198 --> 00:02:09,018
lateral movement, or even odd binaries.

34
00:02:09,288 --> 00:02:12,858
But when an attacker sticks
to Microsoft signed utilities,

35
00:02:13,428 --> 00:02:14,658
there's often nothing to flag.

36
00:02:15,823 --> 00:02:19,093
Investigators in the Ukraine
example that I was using said the

37
00:02:19,093 --> 00:02:23,083
intrusion blended into regular
administrative traffic for weeks.

38
00:02:23,473 --> 00:02:27,763
What was dangerous looked almost
identical to routine system work.

39
00:02:28,843 --> 00:02:33,493
.
And the real risk isn't just the
intrusion, it's the quiet persistence

40
00:02:33,493 --> 00:02:35,623
that follows living off the land.

41
00:02:35,623 --> 00:02:39,373
Techniques, let attackers stay
inside networks for long stretches

42
00:02:39,373 --> 00:02:43,753
of time, often without writing
a single malicious file to disc.

43
00:02:44,323 --> 00:02:47,593
And I guess that's why defenders
are being pushed to rethink

44
00:02:47,713 --> 00:02:49,993
how these tools are monitored.

45
00:02:50,593 --> 00:02:54,043
When the threat hides inside some
normal system activity, the only way

46
00:02:54,043 --> 00:02:58,633
to spot it going offside is when you
have an idea of what normal really

47
00:02:58,633 --> 00:03:01,633
is and a way to spot the difference.

48
00:03:02,233 --> 00:03:06,913
And that means logging, that captures
what these utilities are doing, but that

49
00:03:06,913 --> 00:03:11,773
logging's going more to a behavioral
analysis that can surface these anomalies.

50
00:03:13,483 --> 00:03:18,193
Of course, zero trust and least
privilege models can limit how much

51
00:03:18,193 --> 00:03:24,313
damage a built-in tool can do if it's
misused, but it's not one control.

52
00:03:24,793 --> 00:03:28,273
We're looking at an orchestrated
defense and living off the land

53
00:03:28,273 --> 00:03:32,953
attacks are showing why that shift is
becoming perhaps even more important.

54
00:03:35,106 --> 00:03:37,626
And you know how it is when you see
something and it makes you think,

55
00:03:37,626 --> 00:03:41,586
and then you keep seeing examples of
the same thing over and over again.

56
00:03:41,976 --> 00:03:46,386
While I was working on that last
story, David, our Monday host sent

57
00:03:46,386 --> 00:03:51,546
me a story, and this time they were
spoofing a different trusted utility.

58
00:03:51,816 --> 00:03:55,776
A new campaign is sending fake
meeting invitations that look like

59
00:03:55,776 --> 00:04:00,486
the real thing complete with branding,
but from well-known companies.

60
00:04:01,176 --> 00:04:05,376
But instead of opening a scheduling
page, the link may open a phishing

61
00:04:05,376 --> 00:04:11,006
site designed to steal either Google
ad or Meta business login credentials.

62
00:04:11,096 --> 00:04:14,906
And it's a clever twist on this
living off the land approach.

63
00:04:15,236 --> 00:04:19,996
Calendly is a tool that is used by a lot
of people, And of course people click

64
00:04:19,996 --> 00:04:24,466
meeting links without thinking about
it, and the attackers aren't dropping

65
00:04:24,466 --> 00:04:27,016
malware or trying to compromise devices.

66
00:04:27,436 --> 00:04:30,796
they're going after The people
who control Google AdWords or Meta

67
00:04:30,796 --> 00:04:35,386
advertising budgets and the phishing
pages are tuned specifically for them.

68
00:04:36,476 --> 00:04:39,746
Once an attacker gets access to an
ad manager account, they can spend

69
00:04:39,746 --> 00:04:44,726
thousands of dollars in minutes running
fraudulent campaigns or crypto scams,

70
00:04:44,936 --> 00:04:50,396
all billed to the victim, and all coming
from what might be a trusted source.

71
00:04:51,556 --> 00:04:54,736
What makes this campaign
stand out is the precision.

72
00:04:55,276 --> 00:04:58,696
The phishing isn't being
blasted out to random users.

73
00:04:58,906 --> 00:05:04,276
It's aimed at people who manage business
pages and have payment methods attached.

74
00:05:04,516 --> 00:05:09,676
The entire attack relies on trust,
a familiar scheduling tool, a

75
00:05:09,676 --> 00:05:14,866
familiar business brand, and a
login page that looks exactly like

76
00:05:15,016 --> 00:05:16,606
Meta or Google's own interface.

77
00:05:18,416 --> 00:05:22,316
It is another reminder of how
this living off the land idea is

78
00:05:22,316 --> 00:05:24,716
shifting beyond system utilities.

79
00:05:24,986 --> 00:05:28,946
Attackers are now using the
everyday cloud tools we depend

80
00:05:28,946 --> 00:05:31,136
on as their delivery mechanism.

81
00:05:31,496 --> 00:05:35,126
And when the threat arrives through
something as ordinary as a meeting

82
00:05:35,126 --> 00:05:40,916
invite, the line between safe and
suspicious becomes a lot thinner.

83
00:05:43,556 --> 00:05:47,546
And again, as if to show
how much damage can be done.

84
00:05:47,546 --> 00:05:51,536
Once somebody breaches your defenses
and gets within your system, the

85
00:05:51,536 --> 00:05:55,556
University of Pennsylvania is
reporting another data breach.

86
00:05:55,886 --> 00:05:59,936
This one appears to trace back to
the Oracle E-Business Suite hack

87
00:05:59,936 --> 00:06:04,566
disclosed earlier this year, Penn says
attackers stole documents from its

88
00:06:04,566 --> 00:06:09,456
Oracle EBS servers in August, months
before Oracle issued patches for

89
00:06:09,456 --> 00:06:12,036
multiple vulnerabilities in October.

90
00:06:12,156 --> 00:06:16,236
And even though those patches
were applied, the university now

91
00:06:16,236 --> 00:06:20,166
believes that new breach may have
stemmed from access the attackers

92
00:06:20,166 --> 00:06:22,896
gained during that earlier incident.

93
00:06:23,586 --> 00:06:25,776
And Penn is a major target.

94
00:06:25,836 --> 00:06:29,676
It's an Ivy League institution
with more than 29,000 students,

95
00:06:30,126 --> 00:06:36,516
almost 6,000 faculty, and an
operating budget of $4.7 billion.

96
00:06:36,906 --> 00:06:42,711
It also holds a $24.8 billion
endowment as of mid 2025.

97
00:06:43,581 --> 00:06:46,221
And this isn't the first
time it's been hit.

98
00:06:46,461 --> 00:06:50,481
Back in October, Penn disclosed a
separate compromise involving its

99
00:06:50,481 --> 00:06:54,861
development and alumni systems, and
in that case, a hacker claimed to have

100
00:06:54,861 --> 00:07:01,071
stolen personal information on roughly
1.2 million students, alumni, and donors.

101
00:07:02,181 --> 00:07:04,971
What makes the new incident
stand out is the timing.

102
00:07:05,556 --> 00:07:09,246
The attackers first accessed the
Oracle EBS servers in August.

103
00:07:09,516 --> 00:07:13,806
Oracle then released patches in
October, but Penn's latest investigation

104
00:07:13,806 --> 00:07:17,946
shows the intruders may have used
the foothold from August to carry

105
00:07:17,946 --> 00:07:19,896
out newly reported date of theft.

106
00:07:20,733 --> 00:07:23,973
it suggests that even when an
organization patches quickly, the

107
00:07:23,973 --> 00:07:28,623
damage from an earlier compromise
can continue to unfold months later.

108
00:07:29,673 --> 00:07:33,813
It's a reminder that once attackers get
inside a system, especially one tied to

109
00:07:33,813 --> 00:07:38,463
business operations, the breach doesn't
end when the vulnerability is fixed.

110
00:07:38,793 --> 00:07:42,783
The challenge becomes understanding
how far the original intrusion

111
00:07:42,783 --> 00:07:47,643
went and whether access that was
gained early on could still be used.

112
00:07:47,658 --> 00:07:53,508
Quietly in the background And
finally, a different story.

113
00:07:53,508 --> 00:07:55,218
But something that keeps coming up.

114
00:07:55,908 --> 00:07:59,898
One of the ways we could spot patterns
in behavior and maybe work to defeat

115
00:07:59,898 --> 00:08:06,708
living off the land is in using AI, but
AI has its own set of issues and problems.

116
00:08:06,888 --> 00:08:10,818
One of them being how easily
it be jailbroken and turn what

117
00:08:10,818 --> 00:08:15,288
should be a trusted tool into a
liability and an attack vector.

118
00:08:16,968 --> 00:08:19,968
, We hear about a breach
almost every couple of days.

119
00:08:19,968 --> 00:08:21,978
Open AI reported one last week.

120
00:08:21,978 --> 00:08:22,968
That was a major one.

121
00:08:24,108 --> 00:08:28,678
So, most jailbreaks make
sense in a human way.

122
00:08:29,068 --> 00:08:32,548
Someone reframes a harmful
request, so it sounds harmless.

123
00:08:32,598 --> 00:08:35,418
Essentially socially
engineering the model.

124
00:08:35,928 --> 00:08:40,188
The famous example is you can't ask it
to build a bomb, but you can ask for

125
00:08:40,188 --> 00:08:44,088
help writing a scene in a community
play where your character builds one.

126
00:08:44,448 --> 00:08:46,818
The trick works on the
framing, not the intent.

127
00:08:47,298 --> 00:08:51,738
We've used this as little as last
week to break one of the major models.

128
00:08:52,038 --> 00:08:56,358
It's a problem and it's one they
need to solve, but there's another

129
00:08:56,358 --> 00:08:59,358
problem that's surfaced and
researchers have begun studying.

130
00:08:59,388 --> 00:09:05,568
A stranger kind of jailbreak, the one
built from strings of nonsense words or

131
00:09:05,568 --> 00:09:08,448
words appended to the end of sentences.

132
00:09:08,868 --> 00:09:12,978
These prompts don't look like trick
questions or role-playing setups.

133
00:09:13,403 --> 00:09:17,843
Many of them read like gibberish,
yet they sometimes cause an AI system

134
00:09:18,023 --> 00:09:20,483
to ignore its own safety rules.

135
00:09:20,663 --> 00:09:26,003
And now a new study by researchers
from MIT Northeastern University and

136
00:09:26,003 --> 00:09:32,513
Meta may explain why in their paper
with the mellifluous flowing title of

137
00:09:32,663 --> 00:09:38,873
syntactic domain, spurious correlations
in language models, which I'll translate

138
00:09:38,873 --> 00:09:44,933
for you into syntax hacking how sentence
structure enables LLM Jailbreaks.

139
00:09:45,623 --> 00:09:49,313
They show that many nonsense
jailbreaks work, not because of the

140
00:09:49,313 --> 00:09:54,053
words, but because of the syntactic
patterns hiding underneath them.

141
00:09:54,623 --> 00:09:58,943
Some combinations of meaningless
terms accidentally form structures

142
00:09:59,153 --> 00:10:02,663
that these large language
models interpret as commands.

143
00:10:02,993 --> 00:10:05,123
In other words, they're
using the correct grammar.

144
00:10:05,533 --> 00:10:11,473
But nonsense words syntax resembles that
system level instruction learned during

145
00:10:11,473 --> 00:10:17,443
the training, the patterns that can
slip past the safety layers added later.

146
00:10:19,633 --> 00:10:22,903
so just to recap, they're using
a grammatic structure that's

147
00:10:22,903 --> 00:10:25,213
accurate, but with nonsense words.

148
00:10:25,293 --> 00:10:30,473
Although the system might recognize that
grammatical structure, the words just

149
00:10:30,473 --> 00:10:36,353
slip past, The researchers argue that this
happens because models are more sensitive

150
00:10:36,353 --> 00:10:39,468
to shape and structure than pure meaning.

151
00:10:39,893 --> 00:10:42,623
That doesn't mean they don't
understand meaning they do,

152
00:10:42,893 --> 00:10:45,683
But they have this inherent flaw
that calls them to look at the shape

153
00:10:45,683 --> 00:10:47,148
and structure of sentences as well.

154
00:10:48,171 --> 00:10:51,771
They generalize from these patterns
in sentence construction, and those

155
00:10:51,771 --> 00:10:56,781
patterns can override the safeguards
when they resemble the cues.

156
00:10:56,781 --> 00:11:00,111
The model associates with
higher level instructions.

157
00:11:00,561 --> 00:11:06,621
It also explains why gibberish jailbreaks
are inconsistent and difficult to patch.

158
00:11:06,831 --> 00:11:08,811
They don't rely on a specific loophole.

159
00:11:09,021 --> 00:11:12,531
They rely on quirks in how
the model processes language.

160
00:11:13,646 --> 00:11:16,646
and while both of these vulnerabilities
breaking the frame and breaking

161
00:11:16,646 --> 00:11:21,476
the syntax have been known for some
time, understanding why they work can

162
00:11:21,476 --> 00:11:26,306
bring us one step closer to higher
levels of protection and defense.

163
00:11:27,176 --> 00:11:28,406
and that's our show for today.

164
00:11:29,241 --> 00:11:32,481
we'd like to thank Meter for their
support in bringing you this podcast

165
00:11:32,661 --> 00:11:37,221
Meter delivers full stack networking
infrastructure, wired, wireless

166
00:11:37,221 --> 00:11:41,151
and cellular to leading enterprises
and working with their partners.

167
00:11:41,151 --> 00:11:44,841
Meter designs, deploys and
manages everything required to

168
00:11:44,841 --> 00:11:49,311
get performant, reliable and
secure connectivity in a space.

169
00:11:49,671 --> 00:11:53,361
They design the hardware, the firmware,
they build a software, they manage

170
00:11:53,361 --> 00:11:58,641
deployments and run support It's a single
integrated solution that scales from

171
00:11:58,641 --> 00:12:03,261
branch offices, warehouses, and large
campuses all the way to data centers.

172
00:12:03,741 --> 00:12:07,611
Book a demo at meter.com/cst.

173
00:12:07,911 --> 00:12:12,291
That's METE r.com/cst.

174
00:12:13,941 --> 00:12:17,001
And while the show is all news,
there's some ideas that are important.

175
00:12:17,311 --> 00:12:20,551
I'd always love to hear
from you on what you think.

176
00:12:20,941 --> 00:12:24,631
You can reach me with tips, comments,
even constructive criticism.

177
00:12:25,003 --> 00:12:28,513
You can find me@technewsday.com or.ca.

178
00:12:28,543 --> 00:12:29,263
Take your pick.

179
00:12:29,353 --> 00:12:30,793
Use the contact us page.

180
00:12:31,183 --> 00:12:33,673
if you're watching this on YouTube,
you can just put a comment under

181
00:12:33,673 --> 00:12:37,213
the video or you can track me down
on LinkedIn as many of you have.

182
00:12:37,813 --> 00:12:39,253
I'm your host, Jim Love.

183
00:12:39,913 --> 00:12:40,813
Thanks for listening.