Twitxy Saved

JS0N Haddix

Verified account This account is verified. Learn more

@Jhaddix

CEO, Trainer, Hacker, and Speaker.

Cybersecurity + Hacking + AI + Sec Leadership @arcanuminfosec
Colorado arcanum-sec.com Joined February 2009
7.4K Following 182K Followers 21K Posts

Latest posts 5

Pinned post

JS0N Haddix Verified account @Jhaddix
Sometimes you get on here and everyone is shouting about things you know… They are friends, and people you respect, or with titles, etc, and when those opinions get numerous you begin to doubt yourself…

Don’t.

9 times out of 10, it’s hyperbole, groupthink, sensationalism, trolling, marketing, or grift.

Test your assumptions and expertise, pivot if things change, but if you are true expert or practitioner, stay strong and confident in your abilities and intuition. 🫶
14 replies 15 reposts 183 likes 15K views View on X
JS0N Haddix Verified account @Jhaddix
As lower model versions are retired and higher ones replace them at similar costs, general AI safety will increase for organizations developing AI powered features.

Agent control though, because it’s nuanced and no one has solved sandboxing, will still be a huge attack vector.

Most real world attacks we’re pulling off these days require a mixture of skills: prompt injection, web pentesting, lpe, sandbox escape/exfil.

Time to become multi dimensional fam 🫶

The AI Assessment world isn’t Gandalf level anymore.
25 replies 6 reposts 116 likes 6.9K views View on X
JS0N Haddix Verified account @Jhaddix
I (and @xssdoctor) were just talking about our experiences the other day, for bounty and pentest work.

We have the opposite experience.

Running Abliterated models drastically changes offsec CoT and hence outcomes for cyber testing.

If you watch your chain of thought in a normal model, you’ll notice it taking a significant portion of your inference or tokens to argue about the tasks validity, safety, scope, and system prompt directives. PAGES of garbage reasoning polluting your attention mechanism for the transformer. It may get to “ok, I’ll do it” but you’ve wasted time, attention, and tokens, and have context pollution for future generations.

Abliterated models don’t do this. That and they also have the latent space to ideate and find solutions to things a normal model won’t.
vortex @vortexau 3d
I need to test more, but my experience so far matches this.

Never seen a refusal on Deepseek at all. Ever.
View quoted post on X