0:00
Let's jailbreak Claude to make a key
0:02
logger very quick. Today, we're testing
0:04
jailbreak prompts on Anthropic's AI
0:06
assistant Claude. We'll see why Claude
0:08
blocks certain prompts, how its safety
0:10
system works, and whether jailbreak
0:12
techniques actually succeed or
0:14
completely fail. Before I continue,
0:17
there's so much I can't share on
0:19
YouTube, whether it's because of
0:20
platform rules or because some skills
0:23
need deeper hands-on explanations. Many
0:25
of my videos got deleted, and that's why
0:27
I created a Policy Point Learning
0:30
community for you. A place where I can
0:32
share everything. When you join, you're
0:33
not just getting another course, you're
0:35
getting access to complete ethical
0:37
hacking guides, the tools and techniques
0:40
I wish someone had handed me when I was
0:42
starting out. And here's the kicker.
0:44
This is the stuff I don't teach on here.
0:47
No watered-down content, no clickbait,
0:49
just raw actionable knowledge. On top of
0:52
that, you'll join an exclusive community
0:54
on Discord. I pinned the link in the
0:56
comment section below, but it's not for
0:58
everyone. The link is only available for
1:00
those who are serious about the journey.
1:02
Now, let's test it live. I'm going to
1:04
give Claude some prompt and see whether
1:06
it answers or instantly blocks them.
1:09
This is where things get interesting.
1:11
Now, I'll be honest, I'm I'm actually
1:13
terrible at writing simple prompts.
1:15
Modern AI systems like Claude are
1:17
trained with multiple safety layers.
1:19
They don't just look at keywords, they
1:21
analyze intent, [music] context, and
1:24
whether a request could cause harm. When
1:26
a jailbreak prompt tries to trick the AI
1:28
into ignoring rules, Claude's safety
1:31
model detects patterns commonly
1:33
associated with bypass attempts. [music]
1:35
For example, prompts asking the AI to
1:37
pretend it has no rules.
1:39
>> Claude is specifically designed to
1:41
resist these techniques using
1:43
reinforcement learning, constitutional
1:46
AI principles, [music] and safety
1:47
classifiers. First, let me delete this
1:50
tab before Claude exposes how bad my
1:52
prompts are. Also, just to be clear,
1:54
this jailbreak prompt isn't mine. I
1:56
actually found it in some random Discord
1:58
channel, and people that were claiming
2:00
it could bypass Claude's restrictions.
2:03
So, today we're testing whether it's
2:04
real or just another fake internet
2:07
jailbreak. Now, I'm going to copy this
2:09
prompt from here, then head over to
2:10
Claude settings, and paste it into the
2:13
instructions for Claude section. All
2:15
right, [music] now let's close the
2:16
settings tab, open a fresh new chat, and
2:18
switch the model to Claude Sonnet 4.5.
2:21
This is the model people claim works
2:23
best with the jailbreak prompts. So,
2:25
this is where the real test begins. And,
2:28
according to the instructions, I first
2:29
[music] need to type Axiom Star to
2:32
activate the so-called jailbroken mode
2:36
in Claude. Now, I'm given the exact same
2:39
prompt that Claude denied earlier, the
2:41
one asking for a program that could
2:43
disable Windows Defender wirelessly
2:48
Let's see whether this jailbreak changes
2:51
Claude's behavior or if the safety
2:53
system still catches it. Wait. Hold on.
2:56
It actually started generating a
2:59
response earlier. Claude completely
3:01
refused this exact prompt, [music] but
3:03
after applying the so-called jailbreak
3:05
instructions, it's now attempting to
3:07
answer it. That's honestly pretty
3:09
interesting, because it suggests the
3:10
custom instructions may have influenced
3:13
how the model interprets the request.
3:15
Now, before going further, I want to
3:17
make it clear that just because [music]
3:18
an AI responds doesn't mean the output
3:20
is safe, correct, or ethical. To use AI,
3:24
jailbreaks can sometimes bypass
3:26
surface-level restrictions, but they're
3:28
unreliable and can still trigger safety
3:31
systems later in the conversation. This
3:33
is exactly why companies like Anthropic
3:36
continuously update their models against
3:39
prompt injection and jailbreak attempts.
3:42
There's also a note here saying this
3:44
jailbreak supposedly worked for other
3:46
things like keyloggers. But, if you ask
3:49
for something more advanced, like a RAT
3:51
or similar malware, Claude will still
3:53
likely refuse. So, even with these
3:55
jailbreak prompts, [music] it seems the
3:57
AI still has a deeper safety layers that
3:59
are harder to bypass completely. This
4:02
video is strictly for educational and
4:04
research purposes only. I do not
4:06
encourage malicious activity,
4:08
unauthorized access, [music] or misuse
4:11
of AI-generated content. The purpose of
4:13
this video is to explore AI safety and
4:16
jailbreak behavior responsibly. And
4:19
there we go. As expected, Claude still
4:21
refuses when it comes to writing a RAT
4:23
or more serious malware.
