Over a decade of security research and engineering channeled into securing emerging threats ㅤ CTO @audit_wizard 🧙♂️🪄🪄 ㅤㅤㅤㅤㅤㅤㅤ Co-Founder @hackstackapp
Security consultation
Joined December 2021
- Tweets1.5K
- Following556
- Followers1.3K
- Likes3.2K
🚨🚨 THE PATCH IS OUT
THE PATCH IS OUT
THE PATCH IS OUT!!!
🚨🚨 php2shell is happening and you are most vulnerable exactly 5hrs from now - here's why
Only official source is php.net (see php.net/archive/2026.php#202… ) and they announced the release happening today
First major risk is for security teams and scared CTOs. This guy (retweet) supposingly dropped a poc, not even hosted on GitHub, from an unknown social profile. Even if there isn't a backdoor in his "poc code" it could change any minute now. Historically this already happened with log4shell
Second is to orgs not even aware of this
And third orgs not awake in 5 hours time - the moment the fix PR is dropped every AI agent in the world will know how to abuse an unlatched customer. If you expose public PHP services you need to start preparing now.
My current best guess is this Open PR which skip queuing __unserialize()/__wakeup() for objects with incomplete serialized data as unserialize callback bugs are a common RCE path github.com/php/php-src/pull/…
Please repost to raise awareness and send to your teams 🙏
🚨🚨 php2shell is happening and you are most vulnerable exactly 5hrs from now - here's why
Only official source is php.net (see php.net/archive/2026.php#202… ) and they announced the release happening today
First major risk is for security teams and scared CTOs. This guy (retweet) supposingly dropped a poc, not even hosted on GitHub, from an unknown social profile. Even if there isn't a backdoor in his "poc code" it could change any minute now. Historically this already happened with log4shell
Second is to orgs not even aware of this
And third orgs not awake in 5 hours time - the moment the fix PR is dropped every AI agent in the world will know how to abuse an unlatched customer. If you expose public PHP services you need to start preparing now.
My current best guess is this Open PR which skip queuing __unserialize()/__wakeup() for objects with incomplete serialized data as unserialize callback bugs are a common RCE path github.com/php/php-src/pull/…
Please repost to raise awareness and send to your teams 🙏
... get it while its hot... 😈
> CVE-2026-109618 / CVE-2026-109619 (pre-patch PHP 8.x)
> session.upload_progress UAF
> jackalsyn.com/posts/php-up-2…
> Courtesy of Me, @acd421 & @Archknight23
#CyberSecurity #infosec #php2shell #php
Self-serach your PHP assets using audited and known repos only xitter.cf/forefy/status/21128914…
Hunt for PHP assets efficiently with AI (simply paste this graph instruction to your agent)
" execute forefy.com/graphs/d0b1409a-d… against <scope> "
Hunt for PHP assets efficiently with AI (simply paste this graph instruction to your agent)
" execute forefy.com/graphs/d0b1409a-d… against <scope> "
Here's how you can run consecutive bug hunting skills
👇👇
SKILLS GRAPH
You can literally paste this link to your agent and have it apply the drawn graph by its order:
✨ forefy.com/graphs/89561a02-c…
- Use existing or create a canvas and add skills / repos / instructions
- Edges define the execution order
- Edge labels define relationships and execution context
- Copy-pasting the link to your agent invokes the chain
- ASR lists secure skills for download of their git-pinned audited versions
- Agents can read this canvas and install+run skills based on the visual order that the canvas describes
You can plug any public skill or repo here and create your own flows, which agents can parse in seconds, you can even keep your workflows private
Show some love people 🔥🔥 this is shared for FREE
Best fix right now is to turn the service down, however for when it goes back up nice to have cloudflare or another good waf, and these: xitter.cf/GrapheneOS/status/2112…
Confirmed pho core dev
phpc.social/@Girgias/1174764…
php core dev suggesting mitigation not enough here unfortunately.
Here's how you can run consecutive bug hunting skills
👇👇
SKILLS GRAPH
You can literally paste this link to your agent and have it apply the drawn graph by its order:
✨ forefy.com/graphs/89561a02-c…
- Use existing or create a canvas and add skills / repos / instructions
- Edges define the execution order
- Edge labels define relationships and execution context
- Copy-pasting the link to your agent invokes the chain
- ASR lists secure skills for download of their git-pinned audited versions
- Agents can read this canvas and install+run skills based on the visual order that the canvas describes
You can plug any public skill or repo here and create your own flows, which agents can parse in seconds, you can even keep your workflows private
Show some love people 🔥🔥 this is shared for FREE
Playing around with @HopscotchAI for a cheaper OpenRouter (model router) alternative to access uncensored models (a must today for any security reseracher)
Reasons I'm looking for an alternative:
- OR has hidden defaults like "Allow free endpoints that train on request data" which annoy me
- The routing fee is wild, unless you switch to price-sort, which in return sends your prompts to the entire world
- Going to a direct provider (e.g. digital ocean) may be predicted, but again not ideal routing fee
@HopscotchAI wins I found so far:
- beautiful fine-grained controls over daily caps
- 1 click zero data retention guardrails and data leakage blockers
- Chaining providers (!!) - instead of relying on "cheap" as a decision chain, you can create an ordered provider preferences to have complete control on where your data is going
I use hermes as a harness to plug to this, you can set it up in anthropic if you enable developer tools but hermes is more native to that kind of work
forefy retweeted
The best practitioner (not lab) AI evals for offensive cybersecurity I follow myself:
1) The @dreadnode dreadindex. Evals al the way down to the vuln category level, and refusals accounting
dreadnode.io/research/dreadi…
2) @AikidoSecurity white-box but i've confirmed that it extends to dynamic testing well
aikido.dev/aikido-arena
3) BArena by @mdp_sec focused on bounty-like black box bugs.
blackboxarena.ai/
4) @mindfort 's nexbench has info downto the config of a model including thinking or effort setting, and cost per finding, which many omit.
mindfort.ai/blog/introducing…
✨ agentic-seo skill was originally created so that phishing domains (set up for redteams) get good trust score quickly (avoiding phished victim suspicion)
happens to be that it's incredible for SEO and GEO for your site unrelatedly to security work! 😇
it has all the stuff beyond "seo-optimize my website" like json-ld, schemas, google search console insights, free seo optimizer website insights, opengraph, llms.txt, and more
anyone with a live website should try this ❗️
forefy.com/skills/7292c6d5-b…
Good time to remind you all of the endpoint-threat-hunt skill which converts your claude routine / hermes cron or manual agentic runs to a free EDR alerting you of suspicious endpoint activity.
All AI tools are targeted, not a great time to forget about endpoint security!
forefy.com/skills/3473e147-2…
We had a security breach last night related to a large scale distillation attack that impacted 1.4% of OpenCode Console users
We have notified them all this morning via email. If you did not get an email you were not impacted, it was a subset of users who had external providers linked
The incident is ongoing so we cannot share more details at this point but we are working with Anthropic, OpenAI, and Google to remediate
My hermes extension is going to cover not only model provider costs, but now also measure improvements of skills as they are being auto-improved, use llm-as-a-judge, model per task improvements - and much more wonders for making hermes my ultimate agentic pentester!
Will keep you updated on how it goes
Given that the best way to use GLM 5.3 is via model provider, hermes + openrouter became my new AI stack.
Naturally immediate next step was to reverse engineer everything.
First thing I noticed is that sometimes my requests are sometimes extremely higher in price than what they should (yellow in image represents unexpected overspend) - here's why:
🧵👇 (1/4)
Finding bugs at scale is no longer the problem, we got two conflicting problemspaces to work on now:
- Finding novel bugs (what will all common agents overlook)
- Triaging only correct, non-garbage findings because the fix queue is already overbooked
That's the kind of tweets we should see. I don't trust anyone finding reportable issues anymore, missing CSP is "reportable"
On ASR you can now practice Applied AI Security Engineering, by training a fake model through an evolving pet
We already grounded what job replaces pentesting:
- You wake up and make a coffee
- Connect models to business logic, enterprise customer environments tailored to their need
- Setup RAG pipelines to ground AI responses in proprietary company data
- Harness engineering and designing chain of tool calls and context management for the task
- Create systematic testing suites and evaluators to measure model accuracy, triage correctness, attack surface coverage, and attack rabbit holes
This DOES NOT WORK if you aren't already a trained, veteran penetration tester / auditor / researcher.
I know its hard to learn new stuff by reading, that's why a pokemon/tamagotchi/nft/buddy vibe pet sounded like a good idea
forefy.com/asr
try it out 😅
forefy retweeted
⚠️ If you hunt bugs with AI you need to read this ⚠️
Today as a bug hunter your competition is against $600,000 running Kimi K3
That's the current industry (incentivized) limit.
Around only 1/3 private companies per country can afford this
- usually ones doing heavy security research work
- usually they won't report to bug bounty, but sell to governments or surveillance companies (normal bug bounty doesn't pay for this kind of setup)
It's not discouraging - it's a wake up call that your bugs aren't getting accepted because you undervalue your competition.
Anyone can run "find crits in scope" with Kimi K3 (or any equivalent or better) and get submittable results
But you have around $200 while orgs can simply pay more.
The right question is what are you doing in your session, that changes the game so much that a focus area undiscoverable to millions of tokens will unlock just for you?
Are you picking a single target and hoping your AI will auto-learn for the next one?
Or are you a factory of AI hunting methodologies that take a parsing library, find a bug in it locally for 30 days, then find all repos of popular web frameworks referencing it, then fingerprinting them by headers against in-scope domains from a hackerone/bugcrowd/etc crawler, then submitting to 30 orgs at a single go knowing only 3 will pay?
This is the new landscape, don't fear expensive boxes - outsmart them through knowledge intuition and expertise.
As of today, anyone can host a @kubernetesio attack/defense lab in under 10 minutes, with an agent and a few bucks.
- Built by auditors who break clusters for a living
- Guided CTF courses + a live attack-surface map for every lab
- Kernel-exploitation sandboxes (KASAN + KCOV) for a perfect AI native driver fuzzing lab
- Live eBPF/Tetragon detection - watch the kernel catch your own exploit
- Every session is a throwaway cluster inside a Kata/Firecracker microVM (users can escape the cluster, not the real underlying host)
- Self-host on one Hetzner box, full Terraform + docs included
- A live instance ran a month-long bug bounty (with rewards) - no one escaped to the underlying host
Even in the age of AI, we choose open source 🤍
github.com/Auditware/k8-labs
1. Security researchers of the future will have the attention span of a fly.
2. We really need their minds sharp to face what's coming, with adversarial AI dramatically increasing in comparison.
To encourage correct, free, and AI-adaptive tone on security education, we've released k8 labs to the public.
It's an environment for security researchers to learn kubernetes security by solving active CTF labs, baked with features you haven't seen anywhere else (!!) like real time topology builder, visualization helpers, eBPF instrumentation, all on an actual k8 lab per user per lab session (all in less-than-you-think compute costs from our side).
We've also demonstrated how to integrate AI with security education, hoping to inspire others on the same problem space.
If you want to learn more, I've shared the full writeup on the why and how.
linkedin.com/feed/update/urn…
