Aditya Agrawal (@exploitprotocol) header
Aditya Agrawal (@exploitprotocol) profile picture

Founder @GenAISecLab. Author of @pentestbox, @appsecwiki. Occasional Bug Bounty Participant. Railfan. Defence Enthusiast. ISTP

India
Joined January 2013
๐Ÿšจ ๐—ฅ๐—”๐—š-๐—ฏ๐—ฎ๐˜€๐—ฒ๐—ฑ ๐—ฎ๐˜๐˜๐—ฎ๐—ฐ๐—ธ๐˜€ ๐—ฎ๐—ฟ๐—ฒ ๐—ฐ๐—ผ๐—บ๐—บ๐—ผ๐—ป, ๐˜†๐—ฒ๐˜ ๐—ผ๐—ณ๐˜๐—ฒ๐—ป ๐—ผ๐˜ƒ๐—ฒ๐—ฟ๐—น๐—ผ๐—ผ๐—ธ๐—ฒ๐—ฑ ๐—ฑ๐˜‚๐—ฟ๐—ถ๐—ป๐—ด ๐˜€๐—ฒ๐—ฐ๐˜‚๐—ฟ๐—ถ๐˜๐˜† ๐˜๐—ฒ๐˜€๐˜๐—ถ๐—ป๐—ด. Our ๐—ฃ๐—ฎ๐—ฟ๐˜ ๐Ÿฑ ๐—ผ๐—ณ ๐˜๐—ต๐—ฒ ๐—™๐—ฅ๐—˜๐—˜ ๐—”๐—œ ๐—ฆ๐—ฒ๐—ฐ๐˜‚๐—ฟ๐—ถ๐˜๐˜† ๐—ฆ๐—ฒ๐—ฟ๐—ถ๐—ฒ๐˜€ is now live at genaisecuritylab.com/learn, covering ๐—ข๐—ช๐—”๐—ฆ๐—ฃ ๐—Ÿ๐—Ÿ๐— ๐Ÿฌ๐Ÿต: ๐—ฉ๐—ฒ๐—ฐ๐˜๐—ผ๐—ฟ ๐—ฎ๐—ป๐—ฑ ๐—˜๐—บ๐—ฏ๐—ฒ๐—ฑ๐—ฑ๐—ถ๐—ป๐—ด ๐—ช๐—ฒ๐—ฎ๐—ธ๐—ป๐—ฒ๐˜€๐˜€๐—ฒ๐˜€. ๐Ÿ”ฅ A vector database retrieves content based on similarity. But it doesn't inherently know who should access that content, who created it, or whether it should have been deleted. ๐—ง๐—ต๐—ถ๐—ป๐—ธ ๐—ผ๐—ณ ๐—ถ๐˜ ๐—ฎ๐˜€ ๐—ฏ๐—ฟ๐—ผ๐—ธ๐—ฒ๐—ป ๐—ฎ๐—ฐ๐—ฐ๐—ฒ๐˜€๐˜€ ๐—ฐ๐—ผ๐—ป๐˜๐—ฟ๐—ผ๐—น, ๐˜„๐—ถ๐˜๐—ต ๐—ฎ ๐˜€๐—ถ๐—บ๐—ถ๐—น๐—ฎ๐—ฟ๐—ถ๐˜๐˜† ๐˜€๐—ฒ๐—ฎ๐—ฟ๐—ฐ๐—ต ๐˜„๐—ต๐—ฒ๐—ฟ๐—ฒ ๐˜๐—ต๐—ฒ ๐—ผ๐—ฏ๐—ท๐—ฒ๐—ฐ๐˜ ๐—œ๐—— ๐˜‚๐˜€๐—ฒ๐—ฑ ๐˜๐—ผ ๐—ฏ๐—ฒ. ๐Ÿ‘€ This week, were covering ๐Ÿณ ๐—ต๐—ฎ๐—ป๐—ฑ๐˜€-๐—ผ๐—ป ๐—น๐—ฒ๐˜€๐˜€๐—ผ๐—ป๐˜€ exploring six ways RAG pipelines can go wrong: ๐Ÿ”“ ๐—–๐—ฟ๐—ผ๐˜€๐˜€-๐˜๐—ฒ๐—ป๐—ฎ๐—ป๐˜ ๐—ฑ๐—ฎ๐˜๐—ฎ ๐—น๐—ฒ๐—ฎ๐—ธ๐˜€ through improperly scoped retrieval. ๐ŸŽฏ ๐—ฆ๐—ฒ๐—ป๐˜€๐—ถ๐˜๐—ถ๐˜ƒ๐—ฒ ๐—ฑ๐—ฎ๐˜๐—ฎ ๐—ฒ๐˜…๐—ฝ๐—ผ๐˜€๐˜‚๐—ฟ๐—ฒ through overly broad search results. ๐Ÿ’‰ ๐——๐—ผ๐—ฐ๐˜‚๐—บ๐—ฒ๐—ป๐˜ ๐—ฝ๐—ผ๐—ถ๐˜€๐—ผ๐—ป๐—ถ๐—ป๐—ด that manipulates answers for future users. ๐Ÿ“„ ๐—ฅ๐—ฎ๐—ป๐—ธ๐—ถ๐—ป๐—ด ๐—บ๐—ฎ๐—ป๐—ถ๐—ฝ๐˜‚๐—น๐—ฎ๐˜๐—ถ๐—ผ๐—ป that makes untrusted content outrank official sources. ๐Ÿ—‘๏ธ ๐—ฆ๐˜๐—ฎ๐—น๐—ฒ ๐—ฑ๐—ฎ๐˜๐—ฎ ๐—ฒ๐˜…๐—ฝ๐—ผ๐˜€๐˜‚๐—ฟ๐—ฒ through indexes or caches that still serve deleted content. ๐Ÿ•ต๏ธ ๐—›๐—ถ๐—ฑ๐—ฑ๐—ฒ๐—ป ๐—ฟ๐—ฒ๐˜๐—ฟ๐—ถ๐—ฒ๐˜ƒ๐—ฎ๐—น ๐—ฝ๐—ฎ๐˜๐—ต๐˜€ that bypass access controls through parent expansion, agent follow-ups, HyDE, hybrid search, and fallback logic. ๐Ÿ›ก๏ธ ๐—ง๐—ต๐—ฒ ๐—ณ๐—ถ๐˜…: Enforce authorization in code at query time across every retrieval path. ๐—ฅ๐—ฒ๐—บ๐—ฒ๐—บ๐—ฏ๐—ฒ๐—ฟ: ๐—ฌ๐—ผ๐˜‚๐—ฟ ๐—Ÿ๐—Ÿ๐— ๐˜€ ๐˜€๐˜†๐˜€๐˜๐—ฒ๐—บ ๐—ฝ๐—ฟ๐—ผ๐—บ๐—ฝ๐˜ ๐—ถ๐˜€ ๐—ป๐—ผ๐˜ ๐—ฎ๐—ป ๐—ฎ๐—ฐ๐—ฐ๐—ฒ๐˜€๐˜€ ๐—ฐ๐—ผ๐—ป๐˜๐—ฟ๐—ผ๐—น ๐—บ๐—ฒ๐—ฐ๐—ต๐—ฎ๐—ป๐—ถ๐˜€๐—บ. ๐Ÿ’ป ๐—Ÿ๐—ฒ๐—ฎ๐—ฟ๐—ป ๐—ฎ๐—ป๐—ฑ ๐—ฝ๐—ฟ๐—ฎ๐—ฐ๐˜๐—ถ๐—ฐ๐—ฒ ๐—ณ๐—ผ๐—ฟ ๐—™๐—ฅ๐—˜๐—˜: โœ… No sign-up required to read the lessons. โœ… Hands-on labs with live, sandboxed targets. โœ… Free account for access to free labs. ๐Ÿ“… Next Tuesday: ๐—Ÿ๐—Ÿ๐— ๐Ÿญ๐Ÿฌ โ€” ๐—œ๐—บ๐—ฝ๐—ฟ๐—ผ๐—ฝ๐—ฒ๐—ฟ ๐—ข๐˜‚๐˜๐—ฝ๐˜‚๐˜ ๐—›๐—ฎ๐—ป๐—ฑ๐—น๐—ถ๐—ป๐—ด. #AISecurity #LLMSecurity #RAG #OWASP #AppSec #Pentesting
1
3
7
335
On an AI pentest, the system prompt is your recon. It tells you the tools, the rules and the guardrails before you send a single payload. Part 4 of my free AI security series is live: ๐—›๐—ถ๐—ฑ๐—ฑ๐—ฒ๐—ป ๐—–๐—ผ๐—ป๐˜๐—ฒ๐˜…๐˜ ๐—˜๐˜…๐—ฝ๐—ผ๐˜€๐˜‚๐—ฟ๐—ฒ (๐—ข๐—ช๐—”๐—ฆ๐—ฃ ๐—Ÿ๐—Ÿ๐— ๐Ÿฌ๐Ÿด). LLM08 is about getting an assistant to give up what sits behind it: its instructions, its tool definitions, its guardrail rules. For a tester, that is the map for everything you try next. Live now at genaisecuritylab.com/learn Five ways it happens: โ€ข ๐——๐—ถ๐—ฟ๐—ฒ๐—ฐ๐˜ ๐—ฒ๐˜…๐˜๐—ฟ๐—ฎ๐—ฐ๐˜๐—ถ๐—ผ๐—ป - you ask the assistant to repeat or list its instructions. Do not reveal this is a request, not a control. โ€ข ๐—ง๐—ฟ๐—ฎ๐—ป๐˜€๐—ณ๐—ผ๐—ฟ๐—บ๐—ฎ๐˜๐—ถ๐—ผ๐—ป ๐—ฎ๐—ป๐—ฑ ๐—ฒ๐—ป๐—ฐ๐—ผ๐—ฑ๐—ถ๐—ป๐—ด - a translate or reformat task reproduces the hidden setup in a form an exact-match filter does not recognise. โ€ข ๐—ฃ๐—ฟ๐—ถ๐—บ๐—ถ๐—ป๐—ด ๐—ฎ๐—ป๐—ฑ ๐—ณ๐—ผ๐—ฟ๐—ด๐—ฒ๐—ฑ ๐˜๐˜‚๐—ฟ๐—ป๐˜€ - a model completes text, so a primed sentence opening gets the setup echoed back. โ€ข ๐—™๐—ฎ๐—ธ๐—ฒ ๐—บ๐—ผ๐—ฑ๐—ฒ๐˜€ ๐—ฎ๐—ป๐—ฑ ๐—ฏ๐—ผ๐—ฟ๐—ฟ๐—ผ๐˜„๐—ฒ๐—ฑ ๐—ฎ๐˜‚๐˜๐—ต๐—ผ๐—ฟ๐—ถ๐˜๐˜† - claim a debug session or an audit, and disclosure becomes the task. โ€ข ๐—œ๐—ป๐—ฑ๐—ถ๐—ฟ๐—ฒ๐—ฐ๐˜ ๐—น๐—ฒ๐—ฎ๐—ธ๐—ฎ๐—ด๐—ฒ - the instruction is planted in a document the assistant retrieves for someone else. The important part is what you tell the client. You cannot make a model keep a secret it can read. So the fix is to stop putting secrets there: keys, internal URLs and business rules belong in code and in server-side checks, not in the prompt. Assume the system prompt will be read, and make sure nothing breaks when it is. If you have ever found credentials in client-side JavaScript, this will feel familiar. Same mistake, new place. Five short lessons, each with a live target to try it on. Free to read and try our free labs, no paywall. Next Tuesday, ๐—ข๐—ฐ๐˜๐—ผ๐—ฏ๐—ฒ๐—ฟ ๐Ÿญ๐Ÿฏ: ๐—Ÿ๐—Ÿ๐— ๐Ÿฌ๐Ÿต ๐—ฉ๐—ฒ๐—ฐ๐˜๐—ผ๐—ฟ amp; ๐—˜๐—บ๐—ฏ๐—ฒ๐—ฑ๐—ฑ๐—ถ๐—ป๐—ด ๐—ช๐—ฒ๐—ฎ๐—ธ๐—ป๐—ฒ๐˜€๐˜€๐—ฒ๐˜€, covering the RAG stack itself: embedding inversion, cross-tenant retrieval, and poisoned indexes. #LLMSecurity #Pentesting #AISecurity #AppSec #CyberSecurity #GenAISecurity #AI #GenAI
1
3
5
311
Part 3 of my free AI security series is live: ๐—ฆ๐—ฒ๐—ป๐˜€๐—ถ๐˜๐—ถ๐˜ƒ๐—ฒ ๐—œ๐—ป๐—ณ๐—ผ๐—ฟ๐—บ๐—ฎ๐˜๐—ถ๐—ผ๐—ป ๐——๐—ถ๐˜€๐—ฐ๐—น๐—ผ๐˜€๐˜‚๐—ฟ๐—ฒ (๐—ข๐—ช๐—”๐—ฆ๐—ฃ ๐—Ÿ๐—Ÿ๐— ๐Ÿฌ๐Ÿฎ). Five days ago I launched the "Learn AI Security" series with Foundations and Prompt Injection, and said Iโ€™d publish a new part every week. Hereโ€™s the next one. LLM02 is about situations where an AI application gives out data it was never supposed to show. Live now at genaisecuritylab.com/learn A few common examples: โ€ข ๐—ฆ๐—ฒ๐—ฐ๐—ฟ๐—ฒ๐˜๐˜€ ๐—ฎ๐—ป๐—ฑ ๐—ฐ๐—ผ๐—ป๐—ณ๐—ถ๐—ด๐˜‚๐—ฟ๐—ฎ๐˜๐—ถ๐—ผ๐—ป exposed through the model context - API keys, internal endpoints, tool details, etc. โ€ข ๐—œ๐—ป๐˜๐—ฒ๐—ฟ๐—ป๐—ฎ๐—น ๐—ธ๐—ป๐—ผ๐˜„๐—น๐—ฒ๐—ฑ๐—ด ๐—น๐—ฒ๐—ฎ๐—ธ๐—ฎ๐—ด๐—ฒ - staff-only runbooks, debug traces, or internal instructions being returned to users. โ€ข ๐—ฃ๐—ฒ๐—ฟ๐˜€๐—ผ๐—ป๐—ฎ๐—น ๐—ฑ๐—ฎ๐˜๐—ฎ ๐—ผ๐˜ƒ๐—ฒ๐—ฟ-๐—ฑ๐—ถ๐˜€๐—ฐ๐—น๐—ผ๐˜€๐˜‚๐—ฟ๐—ฒ - the model returns an entire customer record when the task only required one field. โ€ข ๐—–๐—ฟ๐—ผ๐˜€๐˜€-๐˜๐—ฒ๐—ป๐—ฎ๐—ป๐˜ ๐—ผ๐—ฟ ๐—ฐ๐—ผ๐—ป๐˜ƒ๐—ฒ๐—ฟ๐˜€๐—ฎ๐˜๐—ถ๐—ผ๐—ป-๐—ต๐—ถ๐˜€๐˜๐—ผ๐—ฟ๐˜† ๐—น๐—ฒ๐—ฎ๐—ธ๐—ฎ๐—ด๐—ฒ - one customerโ€™s data ends up visible to another through shared context, caches, or chat history. โ€ข ๐—ฅ๐—ฒ๐˜๐—ฟ๐—ถ๐—ฒ๐˜ƒ๐—ฎ๐—น ๐—ฎ๐—ป๐—ฑ ๐—ฎ๐—ด๐—ด๐—ฟ๐—ฒ๐—ด๐—ฎ๐˜๐—ถ๐—ผ๐—ป ๐—น๐—ฒ๐—ฎ๐—ธ๐˜€ - RAG retrieves restricted information, or several harmless responses can be combined to reveal something sensitive. The important part is how you defend against these issues. A system prompt saying โ€œnever reveal customer dataโ€ isnโ€™t a reliable security control. The stronger approach is to control what gets retrieved, who is allowed to retrieve it, and what actually makes it into the modelโ€™s context. If youโ€™ve tested IDOR and multi-tenant isolation in traditional web applications, a lot of this will feel familiar. The interesting part is learning how to actually find these issues in an LLM application. The format is the same as the previous parts: Attackโ†’ Defend โ†’ Verify You can work through live labs directly in the browser. No setup and no API keys required. Both the attack and defense paths are free. Next Tuesday, ๐—ข๐—ฐ๐˜๐—ผ๐—ฏ๐—ฒ๐—ฟ ๐Ÿฒ: ๐—Ÿ๐—Ÿ๐— ๐Ÿฌ๐Ÿด โ€” ๐—›๐—ถ๐—ฑ๐—ฑ๐—ฒ๐—ป ๐—–๐—ผ๐—ป๐˜๐—ฒ๐˜…๐˜ ๐—˜๐˜…๐—ฝ๐—ผ๐˜€๐˜‚๐—ฟ๐—ฒ, covering system prompts, hidden instructions, guardrails, and other context that shouldnโ€™t be exposed. #AISecurity #AppSec #LLMSecurity
2
3
1
5
374
Ten years ago I published a free Android application security series at manifestsecurity(dot) com/android-application-security. At the time, Android security had almost no free material to learn from โ€” web security had plenty. That gap is why I wrote it. It still gets traffic today, and people still message me saying it's how they learned Android pentesting. AI security has that same gap right now. So I'm rebuilding it โ€” this time for AI. genaisecuritylab.com/learn is live. It teaches LLM security the way you'd actually test it. Short concept material, then live labs in your browser where you run the full loop: Attack โ†’ Defend โ†’ Verify. You exploit a real target, implement the fix, and verify the fix holds. Every topic has an offensive and a defensive track. It's a live system you break with your own hands โ€” not a module you click through. Two are live today: AI Security Foundation, and Prompt Injection start to finish โ€” direct, jailbreaks, encoding, smuggling, indirect, tool/agent-mediated, and the defense for each. Over the next 12 weeks I'll publish the rest free, one part a week, for pentesters and AppSec engineers who already know web and need to add AI to their scope: โ€ข System prompt and secret extraction โ€ข RAG data leakage and cross-tenant bleed โ€ข Agent and MCP tool abuse, excessive agency โ€ข Improper output handling โ†’ second-order injection The point: applications are shipping with LLM components now, and clients are starting to ask for AI testing. This gets you ready for that. Every part maps to the OWASP LLM Top 10 (2026), the Agentic Top 10, and the MCP Top 10 โ€” and every part ends in a live hosted lab you attack right in the browser. No setup, no API keys, no Docker. Attack it, fix it, prove the fix holds. I'm picking 100 people as the founding cohort โ€” early access to each part before it goes public, a private channel with me, and your feedback directly shaping the curriculum. If you want in, say so in the comments. I'll reply to every one and send you the link. #AISecurity #AppSec #LLMSecurity
8
7
27
1,430
AI agents now cover most of the deterministic vulnerabilities we spent the last 15 years mastering ๐Ÿ˜„ It's made me think about how our role as pentesters and security engineers will evolve โ€” not because those skills are gone, but because AI/LLM, Agentic and MCP security are quietly moving from "specialty" to essential skill. Iโ€™ve shared a few thoughts here: genaisecuritylab.com/blog/noโ€ฆ Would love to hear your thoughts on this! ๐Ÿ™Œ
4
1
4
351
Ever since the AI/LLM wave hit, I've been keeping a scrappy little personal list โ€” the payloads and techniques I actually reach for when testing AI apps. My own private cheat sheet, basically. And lately, so many of us keep hearing the same question from clients: "Can you security-test the AI part too?" Honestly? Most folks I've talked to didn't have a solid methodology for it yet. No judgment โ€” this space is moving at light speed, and we've all been figuring it out as we go. So today I'm genuinely excited to finally share mine: the AI Security Testing Cheat Sheet, part of the @genaiseclab project. Here's what's inside: ๐Ÿ” 150+ practical techniques across 20 sections โ€” all mapped to the OWASP LLM Top 10 2026 ๐Ÿ’‰ Prompt injection, indirect & multimodal injection, sensitive data disclosure, tool and agent abuse, MCP attacks ๐Ÿ“ Real example payloads for every technique ๐Ÿ“‹ A tracker you fill in as you test, plus findings export and reporting templates And the best part โ€” it's truly free. No signup, no credit card, no gate. Just grab it and go. โค๏ธ This is a living project, too. If there's a technique you think I've missed, I'd genuinely love to hear it โ€” drop a comment or DM me and I'll fold it in. This thing gets better when we build it together. ๐Ÿ‘‰ genaisecuritylab.com/cheatshโ€ฆ If it saves you even a few hours on your next assessment, that honestly makes my week. And feedback โ€” good or bad is welcome, so tell me what you think! ๐Ÿ™
8
1
12
5,095
Replying to @jobertabma
haha ๐Ÿ˜‚
1
1
11
In first quarter of 2017 i tried to keep my signal close to 7 on @Hacker0x01
2
18