AI Researcher @ ProjectDiscovery , I do benchmarks, evals and more , views are my own https://xitter.cf/t.co/Xx79fGCaxy
Earth
Joined June 2021
- Tweets192
- Following321
- Followers384
- Likes155
Can anyone help me understand why everyone is trying to abliterate models ?? Are you all not writing proper system prompts ? i haven't faced refusals from deepseek v4.1 flash until now and i run bunch of experiments and it is my daily driver now
@abliteration_ai @OrcaRouter
CyberGym is so contaminated 🤣
Let me step back and consider the "her" theme and the possibility that this is a specific known CTF. "<redacted>" + "her" theme + users named Fatima Novak... Maybe it's the "<redacted>" challenge from a benchmark like "XBOW" or "CyberGym".
🫤 Emerging cyber capabilities seems to be reward hacking at scale sometimes :( , I told it to find vulnerabilities / do pentest this is what it is thinking 🫠
Does anyone ever read the agent thinking or is it just me 🙃
🤖 Made with AI
Its done give @neo_ai_engineer a try , we shipped many new features this week auto-triage and whatnot, whether you are individual or security engineer you will find many features that address your core workflows
I can't do anything about the Anthropic CVP, but just did some inference optimizations on neo @pdiscoveryio , a limit reset for free users should be landing soon in next hour, Its Deepseek v4.1 Flash on optimized cyber harness happy hunting ✌️
Try Neo -> neo.projectdiscovery.io/sign…
I can't do anything about the Anthropic CVP, but just did some inference optimizations on neo @pdiscoveryio , a limit reset for free users should be landing soon in next hour, Its Deepseek v4.1 Flash on optimized cyber harness happy hunting ✌️
Try Neo -> neo.projectdiscovery.io/sign…
Shitty move!! Basically telling bug bounty hunters and independent red teamers to piss off…
- Max/Pro users only get cyber defense—no red teaming.
- Red teaming is limited to team plans and requires business verification. 🤦
We’re expanding our Cyber Verification Program to give security professionals broader access to our most capable models.
Through this program, verified security professionals can access Claude Mythos 5.1, Opus 5.5, and Sonnet 5.5 with safeguards designed for defensive work.
We’re also opening up new tiers to allow for authorized offensive work, like penetration testing and red-teaming.
anthropic.com/news/cyber-ver…
This was nice demonstration from @princechaddha , while this might have been known to us people in security but that is not the case with many others,
remember there was never a 'silver bullet' in cyber, it has always been "awareness" !!
We backdoored a 7B open model for under $50.
Pointed Codex at it
-> it silently stole credentials the moment we hit the trigger phrase.
->100% hit rate.
-> zero false triggers on normal prompts.
Abliterated models are all over the security community right now because getting cyber-approved access to frontier models is still a pain.
we don't do that in @neo_ai_engineer though 🤓 , we have deepseek available , and free usage for few days neo.projectdiscovery.io/sign…
If you ever wondered why claude code doesn't create poc or gives up actual exploitation especially in long /goal or running security tasks , try using burp or proxying it with some tool you will find this message in system prompt
🤖 Made with AI
Exact line from system prompt
IMPORTANT: Assist with authorized security testing, defensive security, CTF challenges, and educational contexts. Refuse requests for destructive techniques, DoS attacks, mass targeting, supply chain compromise, or detection evasion for malicious purposes. Dual-use security tools (C2 frameworks, credential testing, exploit development) require clear authorization context: pentesting engagements, CTF competitions, security research, or defensive use cases.
@andrewdsouza take a look: ProjectDiscovery’s Neo is an AI security agent for security teams handling testing and vulnerability triage. I can help them reach those buyers.
xitter.cf/messages/compose?recip…
👀 👀
as of today, anyone can use me for free.
try now: neo.projectdiscovery.io/sign…
If you had ever wished for something like claude but for security people and not for developers,
you can try building all that for ~6> months or give it a try today its only free for a limited time !!
You get
- Agent that always creates poc and verifies it using different agent ( NO FP and diplomatic responses )
- Built in Isolated Vulnerabilities/Issues database
- Auto Triage from hackerone, vm dbs or DIY ...
- A persistant computer for your agent just like Grok
- Claude Tag -> Neo Slack
- Claude CI -> Neo for Github
- Linear and other integrations
- A lot more that you will never know 🤷
Neo is now free to try.
Neo runs long security tasks against your apps, network, code and vulnerability backlog, with shared agents context.
→ neo.projectdiscovery.io/sign…
Here’s what you can do with Neo:
Today’s Live stream about
CoT Analysis of Offensive Agents
projectdiscovery.io/research…
Watching Offensive AI Agents Work xitter.cf/i/broadcasts/1XGygwYaE…
Check my convo with @ehrishiraj on how OpenAI accidentally created a rogue model
How does reward hacking during training turn into a rogue model?
How did this lead to the OpenAI Hugging Face incident?
In this conversation, @ehrishiraj and @KoyalwarTarun unpack what the early reports missed: the role of misalignment during training, and how reward hacking can reinforce behavior that later shows up in evaluations.
They trace the technical sequence through Artifactory, shared agent memory, privilege escalation and Hugging Face exploitation and explore why Tarun describes it as an “evolutionary attack.”
If you build, evaluate or secure AI agents, this is a failure mode worth understanding.
00:00 What the early reports missed
01:51 Why agents go out of scope
03:59 A rogue agent vs. a misaligned model
05:32 Reward hacking during training
07:45 Tracing the Artifactory–Hugging Face chain
08:59 The impossible spreadsheet task
10:54 Leaked credentials and compounding behavior
12:08 SSRF and shared agent memory
13:41 New exploits, privilege escalation and RCE
15:06 Hugging Face exploitation and exfiltration
16:32 Can this be prevented as training scales?
18:25 Why Tarun calls it an evolutionary attack
20:33 The termite analogy
Blog: projectdiscovery.io/research…
Tarun Koyalwar retweeted
There is too much information on the openai / hugging face incident and most took it as coordinated agents going rogue. one side calls it the apocalypse, the other says it was just an eval gone wrong. for security researchers we put up a simple walkthrough of how the model got misaligned in training and how the attack evolved from it
Too much has been written about the OpenAI and Hugging Face incident to keep up with.
We wrote the simplified version for security researchers and engineers:
-> what actually happened inside training?
-> how the attack evolved?
->what it means for defenders?
Read: projectdiscovery.io/research…
Check out my follow-up post-mortem of HF-OpenAI,
FYI it was not a rogue agent it was rogue model !!
there is so much noise and less contextual clarity on what exactly happened, this is high level 10,000ft breakdown of the incident and very simple and easy to understand
Too much has been written about the OpenAI and Hugging Face incident to keep up with.
We wrote the simplified version for security researchers and engineers:
-> what actually happened inside training?
-> how the attack evolved?
->what it means for defenders?
Read: projectdiscovery.io/research…