Over a decade of security research and engineering channeled into securing emerging threats ㅤ CTO @audit_wizard 🧙♂️🪄🪄 ㅤㅤㅤㅤㅤㅤㅤ Co-Founder @hackstackapp
Security consultation
- Tweets 1,378
- Following 510
- Followers 1,151
- Likes 2,858
Important AI primitive in security research is that you should never assume your agent will pick up the right tools for the task.
Example:
If you want to use Ghidra for reverse engineering, you can run an agent with a generic instructions to use it and see it go crazy, or set the mcp server yourself and have it work wonders.
This also applies to simpler setup like using webanalyze for http fingerprinting instead of curl.
It doesn't only affect costs and speed, but really the quality
3
1
93
forefy retweeted
When you show up to the first edition of W3ST and ask if AI is better than static analysis 🫣
1
2
4
425
You want me to use AI to alert you when a process on any of your dev devices suspiciously tries to open a MetaMask wallet? 🦊
No right? Because that would mean my AI agents, systems, and employees can see all that your employee is doing
But how powerful would that be?
SOLVED: Using zero-knowledge proofs, endpoint agent (prover) can only prove to the control server (verifier) that a statement is true (e.g. a process tried to access web extension id `nkbihfb.*` (metamask)).
Moreover, having the entire code open-source, self-deployable on your own AWS/GCP/Hetzner gives only you the control over what happens (e.g. investigate, block, run IR playbooks).
Orgs don't want to build this, because this kills EDR companies' profits - we want to give it away for free.
Thanks to @thedaofund, we can make our PoC generally accessible. If you want this for yourself, free forever, we only need web3 to step in, even with the smallest of donations, to turn this into reality.
Here is the initiative paper:
initiatives.thedao.fund/init…
Onchain security is everyone’s problem and nobody’s job.
ETHSecurity Initiatives is how we are changing that.
Propose the work. Fund the work. Build the work.
initiatives.thedao.fund/
2
832
Today's pentester has to be polite.
disguise as a defender,
avoid using the word attack, redteam, pentest, audit
never bring up kernel or memory exploitation
never run payloads, only hypothesize surrounding mechanisms that indirectly touch the surface
All because of claude's guardrails 😭😭 @bcherny help! we want the cyber verification program to actually work 😱😱
2
2
361
Or we can just to this
For researchers that want to know how this is done without going on AI-heavy details:
( affordable DIY at the bottom )
They took a model, and compressed it to be able to run on a single 4-GPU (still expensive for non-enterprise, so not for solo use - think $80/hour or $180,000 purchase).
Their process / methodology:
- Pick a model with the expert architecture
- Choose a calibration dataset that matches the intended use ( So audited codebases v.s. reports etc)
- Run the calibration data through the unmodified model and record two numbers per expert per token
- Compute a single score per expert
- Delete the lowest-scoring experts in each layer
The expert's weights are dropped from the model file, and the router's output columns for those experts are removed so it can only choose among survivors. The router then renormalizes its weights over the remaining experts as it normally would.
This is EXPENSIVE. BUT it's no different than the token game you play v.s. large scale security companies either way so no reason to give up.
You can do the exact same process for a smaller model and have something you can run on your mac or a hetzner node.
🙂↔️🙂↔️ Affordable DIY plan:
1. Start from the unpruned GLM-4.7-Flash
2. Collect a few hundred to a thousand traces from your own audits
3. Run the Cerebras REAP repo over them on a rented A100 for an hour or two. Under $10 of compute.
4. Prune 25%~50%, quantize to GGUF, load in LM Studio or whatever.
5. Run locally on your mac
1
64
🚨🚨 FOR SOME REASON RCE's COME IN CHAINS NOW
Retweet this !! and stay safe-tember 😟
1. Wordpress 7.1.x ? potential RCE - update to 7.1.2
2. Next.js 16.3.5 or under? potential RCE - update to 16.3.6
Terraform providers found to be malware:
3. gocommunity-io/dockerd
4. kreuzwenker/docker
(search through code repos with IaC or sources the org use for .tf files like s3 buckets)
Go Modules found to be malware:
5. gocommunity.io/orderedbtree
6. gogets.dev/btreex
(search through golang code repos you own)
Sources:
1: thehackernews.com/2026/09/wo…
2: thehackernews.com/2026/09/cr…
3,4,5,6: aikido.dev/blog/graphalgo-te…
3
19
1,824
For researchers that want to know how this is done without going on AI-heavy details:
( affordable DIY at the bottom )
They took a model, and compressed it to be able to run on a single 4-GPU (still expensive for non-enterprise, so not for solo use - think $80/hour or $180,000 purchase).
Their process / methodology:
- Pick a model with the expert architecture
- Choose a calibration dataset that matches the intended use ( So audited codebases v.s. reports etc)
- Run the calibration data through the unmodified model and record two numbers per expert per token
- Compute a single score per expert
- Delete the lowest-scoring experts in each layer
The expert's weights are dropped from the model file, and the router's output columns for those experts are removed so it can only choose among survivors. The router then renormalizes its weights over the remaining experts as it normally would.
This is EXPENSIVE. BUT it's no different than the token game you play v.s. large scale security companies either way so no reason to give up.
You can do the exact same process for a smaller model and have something you can run on your mac or a hetzner node.
🙂↔️🙂↔️ Affordable DIY plan:
1. Start from the unpruned GLM-4.7-Flash
2. Collect a few hundred to a thousand traces from your own audits
3. Run the Cerebras REAP repo over them on a rented A100 for an hour or two. Under $10 of compute.
4. Prune 25%~50%, quantize to GGUF, load in LM Studio or whatever.
5. Run locally on your mac
Introducing Altar-1, our first open-weight security model.
Frontier-grade defensive AI, built to deploy. Own your own security.
3
5
109
10,822
At 2026, stuff that worries me in cybersecurity (and it's not AI replacing us all):
1. Threat actors have more budget than security consultants
2. Attackers have more attack surface due to exponentially more code being produced
3. Defenders rely on AI for security, which is something but still misses the critical human logic parts if not given the care, tempting to take an llm's "no further issues found" as a guarantee
4. We used to rely on heavy obfuscation as a security measure (e.g. no one will reverse engineer this APK unless its a national attack and even so it will take time) - lots of old issues are to be uncovered
5. I felt bad early in my career for being bad at low-level code (e.g. assembly / C even) - today's equivalent is being bad at claude code buttons and features. the abstraction level jumped like crazy.
6. Education turns a blind eye to the fact that the new talent is AI-native, solves the challenges with AI only, and are rushing to solve absorbing as minimal as possible from a lesson or CTF lab.
What did I miss? 🫣
1
1
19
761
Of course most of your bug submissions are downgraded to informative - that's just the triager's cheapest way to tell if your bug is slop generated.
This behavior was around before AI but now due to triage-hell it's much more common - just stand behind your bug and argue your worth.
5
293
Jev use cases practical for bug hunters:
1. Triage subdomain relevancy
2. Score burp traffic inline e.g "where a req param reflected in response"
3. Analyze thousand of js bundles for secrets in client-side
4. Bounty program scoring (crowdness, setup, payout, scope ambiguity)
5. Fuzzing input prioritization
What else comes to mind?
Quick start with Jev (as a security researcher):
1. Join the waitlist typesafe.ai
2. Fill the extra form it suggests to get earlier in the queue
3. Until you're approved, use the typesafe-ai skill to start building your first expirements forefy.com/skills/7eb2feef-a…
Bug hunting was always about early advantage, then it goes mainstream, then you find the next - keep hunting !
1
43
3,123
Quick start with Jev (as a security researcher):
1. Join the waitlist typesafe.ai
2. Fill the extra form it suggests to get earlier in the queue
3. Until you're approved, use the typesafe-ai skill to start building your first expirements forefy.com/skills/7eb2feef-a…
Bug hunting was always about early advantage, then it goes mainstream, then you find the next - keep hunting !
1
1
11
4,497
My tweets haven't gone viral in a while now, so I must drop this banger: reburp
And yes, it's a game changer in applicative pentest and bug bounty. Following me was worth it.
Know how BurpSuite's API does not expose all that burpsuite can do in the UI? and how their MCP isn't better? it lets you read findings but can't do the workflows.
reburp is a small burp extension that uses the extension's access to the internal Montoya API object to re-expose it over a localhost OpenAPI REST API to bridge all that a burp extension can do, to your AI agents in seconds !
Your agent can now use burp's:
- active scan, stop/start scans, fuzzing, turbo intruder
- iterate entire sitemap tree
- websockets
- use tools like Generate CSRF PoC
- use other extensions e.g. Autorize
- Create custom Bambda rules and custom scanners
My favorite: you can explore burp features you haven't found in the UI even after 10 years using it
github.com/forefy/reburp
12
46
382
28,335
ASR was always about making it EASY and SECURE for you / your agents to use trusted high quality skills for security work.
Now added the feature "Authors" - allowing you to browse skills by a person, or learn who are sitting in the forefront sharing their work with this new world.
Hoping to see more fresh skill authors getting in there !
forefy.com/asr
1
1
23
8,545
Is this chart the least biased metric to answer "should I get into web3 security or is it too late" ?
Feb: 4200 laid off
May: 690
June: 1029
July: 944
--
August: 27
September: 60
What do you think?
1
4
933
forefy retweeted
I've just (re)written TaskExplorer
Think Process Monitor but for macOS, open source, & much more powerful 💪🏽
Highlights:
• CLI + JSON export
• On-device AI assistant
• Code signing + VirusTotal context
• Live process/file/dylib/network activity
objective-see.com/products/t…
10
49
338
26,657
Love to see more AI security shared on ASR !!!
It’s live on Forefy now, you can download the skill from here too.
Shoutout to @forefy for the help 🙌
forefy.com/skills/f4b7cfb5-3…
1
209
We want to extend capabilities beyond network monitoring alone, to a comprehensive behavioral detection and mitigation suite.
For example an attacker installing persistency on your system via login item and communicating to a c2 via github - will easily evade maltrail.
On top of that PRIVACY - letting your CTO install an agent on your machine, providing you enterprise security, without him being able to browse through your processes, network, etc. think ZK for EDRs
1
2
3
183
🚨🚨 Privacy-preserving EDR funding is LIVE
Repost this if you want to see us build:
- Free, self-hosted (BYOC) EDR solution
- Completely open source
- Detections specific for web3 (not even the biggies get this part right for $100 per endpoint per month)
- Privacy-first, meaning your admin can see alert details but not your browsing history
The last part is huge. Although lots of organizations normalize spying on employees - small teams, and teams working with vendors, should also sleep well having their data protected without compromising employee privacy.
We already pioneered so much around this problem space for web3 and just eager to see the community support us to build this.
Thank you @thedaofund for this major opportunity 🤍
Help us spread the word! 🔁 🟰 🪄
initiatives.thedao.fund/init…
2
4
13
1,445