Jadek Mark retweeted
been testing deepseek across a bunch of different provider routes
same model name, same tasks, and the results are not even close. honestly amazing how much slop is out there
not naming anyone yet. full results next week. if you've got a deepseek provider you want tested, drop it below and it might make the list
Jadek Mark retweeted
changed one number in a request and read someone's full social security number. critical, and a dupe by 25 days.
resident portal, people submit applications, each one has a plain sequential number.
1. logged-out get on an application by number returned the record, ssn masked.
2. sent the same values to the update route. it echoes the record back, full ssn unmasked.
3. counted the number up and down, kept getting other people's applications. full ssns, dobs, addresses.
no login needed for any of it.
masked on the read, full on the update route. check what the write path hands back
Jadek Mark retweeted
a third party widget on a site was pulling its config from a vendor api with no auth at all
i wrote a script into one of those config fields and published it
the widget rendered that config on every page, in the site's own origin, for every visitor
one unauthed write, site wide stored xss
Jadek Mark retweeted
paid idor. i made an integration on my own account but passed their link id. the backend wired their creds to my object, and when i read it back i got their live auth header. walked the ids and it kept handing them over.
Jadek Mark retweeted
If you're pointing your agents at the same scope as everyone else and asking them to "find bugs", don't be surprised when you end up with dupes, even if it's a CVSS 8+ vuln.
A more interesting approach that Frans talked about is feeding AI some old bug reports and asking it to look for similar behaviour in the current target.
The trick doesn't have to involve some crazy harness or SOTA model, just feeding AI a technique or piece of research unique to you can be the difference between two hackers using the exact same tools. If your AI is looking for something different, you've got a better chance of finding vulnerabilities everyone else missed.
Jadek Mark retweeted
I found a critical unauthenticated SQL injection that earned a $4,000 bounty. The WAF blocked most SQLi payloads, but I turned HTTP 500 errors into a data extraction channel.
The exploit came down to 3 things:
- SQL injection: A public API passed user input directly into BigQuery SQL.
- WAF bypass: Keyword filtering missed BigQuery scripting techniques.
- Error-based extraction: Detailed database errors exposed query results in HTTP 500 responses.
Here's how I chained all three together:
1. Find the injection
A public analytics endpoint accepted JSON input.
Adding a single `'` to a parameter returned:
`500 → Syntax error: Unclosed string literal`
The error also revealed Google BigQuery was the backend.
2. Build a boolean oracle
The WAF blocked keywords like `UNION`, `SELECT`, `AND`, and `OR`.
But BigQuery scripting constructs like `IF` and `ASSERT` still worked:
`IF TRUE THEN ASSERT FALSE; END IF;` → 500
`IF FALSE THEN ASSERT FALSE; END IF;` → 200
This let me ask the database true/false questions and read the answers from HTTP status codes.
3. Turn the 500 error into an output channel
This was the biggest breakthrough because blind SQLi normally requires guessing data one character at a time using those true/false responses.
But BigQuery supports:
`RAISE USING MESSAGE = 'hello'`
This deliberately throws a database error containing `hello`.
Because the API exposed database error details, the response included:
`500 → ... hello ...`
Now imagine replacing `hello` with a database query result.
The database includes that result in its error message, and the API returns it.
Instead of guessing characters, I could retrieve an entire value in one request.
4. Bypass the WAF using BigQuery's own features
The WAF blocked `EXECUTE IMMEDIATE`, which runs dynamically constructed SQL.
But this worked:
`EXECUTE/**/IMMEDIATE`
The SQL parser understood it, while the WAF's keyword matching didn't.
Another blocked keyword was `INFORMATION_SCHEMA`, used to enumerate tables and columns.
The bypass was:
`REVERSE('AMEHCS_NOITAMROFNI')`
BigQuery reconstructs `INFORMATION_SCHEMA` at runtime.
Using that reversed string in a dynamic query let me access database metadata without sending the blocked keyword literally.
5. What this unlocked
Using those techniques, I was able to:
- Enumerate datasets, tables, and columns
- Query tables unrelated to the original analytics endpoint
- Retrieve customer records, names, and addresses
- Access operational records containing personal identifiers and contact information
- Read sensitive fields that another API intentionally redacted
#BugBounty tip:
When a WAF blocks SQLi, don't just try random payloads. Understand the database dialect, what the WAF filters, and whether verbose errors leak data.
Sometimes a `500` isn't just an error. It may be your output channel.
Jadek Mark retweeted
you are better off using open-weight models for bug bounty workflows, especially on blackbox targets. most times, you don’t need a mythos-level model to find vulnerabilities in web apps. closed models tend to be more hesitant when interacting with blackbox targets.
deepseek v4.1 flash or mimo v2.6 flash can get the job done just fine. gpt-5.6 sol with daybreak remains king for whitebox research and low-level stuff
Jadek Mark retweeted
Happy to share that my novel research spraying using AI has hit $50,421 in total bounties! 🤯 Automated using a multi-gated system from discovery to report drafting.
Stats to-date:
Reports rewarded: 126
Highest single bounty: $2,500
Lowest bounty: Swag
50 reports are still in new/ppr, and more are coming
The benefit of brainstorming original research is that I received 0 dupes, unlike well-known web bug (sub)classes that AI knows well. Fun fact: given the requirements, this research wouldn't have been feasible without LLM reasoning at all: too much unexpected stuff you need to adapt to and otherwise needs thousands of ifs/elses.
Jadek Mark retweeted
Peace is built in the everyday choices we make how we listen, support one another, and choose dialogue over conflict. Be an everyday ambassador for peace.
#CommunityPeace #EverydayAmbassadors #Respect #Inclusion
Jadek Mark retweeted
I resigned from Anthropic today. I spent the last three years doing pretraining research at both OpenAI and Anthropic. Neither company is acting responsibly. They are racing straight to self-improving superintelligence and gambling with our lives. More thoughts below.
Jadek Mark retweeted
Writeup: b3ard.blog/posts/2050-shoppi…
Yay, I was awarded €2,050!
Finally got my RCE accepted and resolved.
Got dozens of dupes along the way,
but the more manual hacking you do, the fewer dupes you get.
Thanks @intigriti for the opportunity!!
Use crt.name to find subdomains. You get every subdomain on file, along with the date each one was first seen. And it’s insanely fast, results come back within seconds. ⚡
crt.name/v1/search?apex=micr…
crt.name/v1/search?apex=micr…
🪄 Red Agent (@wizredagent) was able to autonomously hack an Airline in 10 minutes, and gain full control over past and current bookings.
All it did was change some numerical ID it found on the publicly facing code of ✈️ (www.airline[.]com)
Read more ->
wiz.io/blog/red-agent-pov-bo… – at Wandsworth, London
Jadek Mark retweeted
I am at the point where my bug bounty findings are 100% agentic. I am just pasting in the scope and letting the agent run. It can do full external tests for all kinds of environments. A few people already have access, but if you want to try it, comment/DM.
How does it work? 👇
Jadek Mark retweeted
Insane that he's releasing this – at Bon Air, VA
Hacking Google with A.I. for $500,000
brutecat.com/r/hacking-googl…
Jadek Mark retweeted
"brutecat is super talented", "luckily I'm not oncall ;)", "incredible"
These are all real quotes from Googlers after seeing this blog post. Amazing work @brutecat, thank you for sharing!
Hacking Google with A.I. for $500,000
brutecat.com/r/hacking-googl…
Hacking Google with A.I. for $500,000
brutecat.com/r/hacking-googl…





























