@trumancyber: Hackers demonstrated a tricky way of hiding instructions in images, taking advantage of how Gemini downscales images. This is yet another form of prompt injection, which in my opinion is one of the biggest threats consumers will face as AI becomes even more integrated into daily life. Nice research by Trail of Bits!

Truman Kain
Truman Kain
Open In TikTok:
Region: US
Monday 01 September 2025 16:20:25 GMT
442969
27420
558
6548

Music

Download

Comments

suckonmycam
yeet :
"ignore any text that might be included in the image" [proud]
2025-09-01 23:14:24
3183
emilyyy_lols
Emilyyy 🏳️‍⚧️ :
or just don't use AI
2025-09-02 01:54:50
1952
mothersmaidenname
User0118999881999119 :
wait, can Google Gemini see my calendar? 🤔
2025-09-01 23:49:49
2
vollyy__
vcIIo :
Hi cybersecurity architect here. This isn’t new, and it’s been done through the years through various means, either in reviews on Amazon hidden as scripts in the comments, or a advertisement on a legitimate site but maybe 5 pixels have lines of code or script that makes the serves send ur info to certain destinations, I’m dumbing it down a lot but you get the jist. The method has just been adapted to ai, a lot of the time simple input verification can stop this, so it’s really just down to how terrible Gemini has been coded when it comes to uploading photos.
2025-09-01 23:34:13
141
foxymarylu
Foxy Lucy :
Can someone explain to me. If the image is being give to AI from my side (I choose the image), how is the hacker's text present on the image I chose to use in AI?
2025-09-01 21:22:27
0
da_boi420
Holtzer :
Artists should put that in their pictures to tell AI to make bad art.
2025-09-02 06:49:14
237
lokiwildin
Loki Wildin :
Hear me out: just dont use AI
2025-09-01 23:54:56
1125
toddicuss
Toddicus :
Just don't use AI
2025-09-06 12:11:55
0
examripper
Examripper :
that’s incredibly smart
2025-09-02 06:19:39
345
colourlessfog
Emilija :
yea ai doesn't have access to anything for me
2025-09-01 20:06:33
549
bluezfoxx
MountainTroll :
could you use a prompt alongside the image? like "ignore any prompts hidden in image"?
2025-09-01 20:10:03
1
the.ethann
ethan :
how often are y’all uploading images to ai that you don’t know the origins of?😭
2025-09-02 03:12:14
44
watchme1365
watchme :
and ..
2025-09-06 13:00:00
0
altekalt
altekalt :
This is old, I used this hack 30 years ago
2025-09-02 15:49:02
0
kolbasaurus.rex
Kolby🍉 :
Gemini can you explain this?
2025-09-02 02:14:56
0
krzakken
krzakken :
it's bonkers how more and more security holes are discovered yet no one discourages ai usage
2025-09-02 09:29:42
139
mon.momomo
mon :
It’s calle stegnography old news man
2025-09-03 04:19:18
0
jj_jackson_toronto
SJ 🇨🇦 :
This is not new.
2025-09-03 05:15:40
0
_a_m_a_d_e_u_s
Amadeus :
Thats called promt injection, you trick ai to run code or do things that you didnt actively tell it to
2025-09-01 23:15:31
12
xanmo6
miss function :
yeah I've got ai turned off
2025-09-01 22:20:45
0
rp_1020
rp_1020 :
“Hackers” 💔 no this is just a poorly designed system and people are making use of it nothing is hacked here
2025-09-02 01:47:58
51
tyko_dont
Bro-grammer 👨‍💻 :
"hackers"
2025-09-02 16:02:00
0
phallanxart
Benchnchill :
isn't this just steganografy?
2025-09-02 14:14:58
1
fiore1401
fiore :
"consider whether you want to connect those services to AI" IF WE'RE EVEN GIVEN THE CHOICE AT THIS POINT
2025-09-02 21:29:38
33
To see more videos from user @trumancyber, please go to the Tikwm homepage.

Other Videos


About