HackTricks-wiki · carlospolop · Aug 5, 2026 · Aug 5, 2026
diff --git a/src/AI/AI-Assisted-Fuzzing-and-Vulnerability-Discovery.md b/src/AI/AI-Assisted-Fuzzing-and-Vulnerability-Discovery.md
@@ -3,7 +3,7 @@
 {{#include ../banners/hacktricks-training.md}}
 
 ## Overview
-Large-language models (LLMs) can super-charge traditional vulnerability-research pipelines by generating semantically rich inputs, evolving grammars, reasoning over crash data, and even proposing multi-bug patches.  This page collects the most effective patterns observed during DARPA’s AI Cyber Challenge (AIxCC) finals and other public research.
+Large-language models (LLMs) can super-charge traditional vulnerability-research pipelines by generating semantically rich inputs, evolving grammars, reasoning over crash data, and even proposing multi-bug patches.  This page collects the most effective patterns observed during DARPA’s AI Cyber Challenge (AIxCC) finals and other public research.<sup>[[4]](#references)[[5]](#references)</sup>
 
 What follows is not a description of one specific competition system, but an abstraction of the techniques so you can reproduce them in your own workflows.
 
@@ -135,7 +135,7 @@ Implement a queue where confirmed PoV-validated patches and *speculative* patche
 
 ## 6. Deterministic File-by-File AI Code Review
 
-A frequent failure mode in AI-assisted review is asking one agent to inspect a whole repository and hoping it chooses the right files and grep terms. A more reliable pattern is to **force repository coverage**:
+A frequent failure mode in AI-assisted review is asking one agent to inspect a whole repository and hoping it chooses the right files and grep terms. A more reliable pattern is to **force repository coverage**:<sup>[[1]](#references)</sup>
 
 1. Enumerate source files.
 2. Send **one file at a time** plus minimal context (entrypoint, imports, nearby routes/callers).
@@ -152,8 +152,8 @@ This is token-heavy and noisy, but it is very effective at surfacing **simple hi
 
 ### Two practical bug classes this method finds well
 
-- **PHP controller-dispatch LFI/RCE**: if a request-controlled controller name is concatenated into `require_once()` with no allowlist/path normalisation, traversal sequences such as `../` can make PHP include an unintended local `.php` file. If the attacker can point the include at a planted or otherwise useful PHP file, the LFI becomes code execution. See [File Inclusion / Path Traversal](../pentesting-web/file-inclusion/README.md).
-- **Authenticated command injection in hosting/admin panels**: if account-management fields such as usernames are embedded into shell commands, a low-privileged authenticated user may turn a normal create/delete action into RCE. The impact increases when the panel executes the command as a more privileged service account. See [Command Injection](../pentesting-web/command-injection.md).
+- **PHP controller-dispatch LFI/RCE**: if a request-controlled controller name is concatenated into `require_once()` with no allowlist/path normalisation, traversal sequences such as `../` can make PHP include an unintended local `.php` file. If the attacker can point the include at a planted or otherwise useful PHP file, the LFI becomes code execution. See [File Inclusion / Path Traversal](../pentesting-web/file-inclusion/README.md).<sup>[[1]](#references)</sup>
+- **Authenticated command injection in hosting/admin panels**: if account-management fields such as usernames are embedded into shell commands, a low-privileged authenticated user may turn a normal create/delete action into RCE. The impact increases when the panel executes the command as a more privileged service account. See [Command Injection](../pentesting-web/command-injection.md).<sup>[[1]](#references)</sup>
 
 ### Practical review notes
 
@@ -183,9 +183,11 @@ graph TD
 ---
 
 ## References
-* [Project Black - Local AI for Cyber Security: Finding phpIPAM LFI and myVesta Authenticated RCE](https://projectblack.io/blog/local-ai-for-cyber-security)
-* [Strix](https://github.com/usestrix/strix)
-* [GitHub Copilot community security-review skill](https://github.com/github/awesome-copilot/blob/main/skills/security-review/SKILL.md)
-* [Trail of Bits – AIxCC finals: Tale of the tape](https://blog.trailofbits.com/2025/08/07/aixcc-finals-tale-of-the-tape/)
-* [CTF Radiooo AIxCC finalist interviews](https://www.youtube.com/@ctfradiooo)
+
+- [1] [Project Black - Local AI for Cyber Security: Finding phpIPAM LFI and myVesta Authenticated RCE](https://projectblack.io/blog/local-ai-for-cyber-security)
+- [2] [Strix](https://github.com/usestrix/strix)
+- [3] [GitHub Copilot community security-review skill](https://github.com/github/awesome-copilot/blob/main/skills/security-review/SKILL.md)
+- [4] [Trail of Bits – AIxCC finals: Tale of the tape](https://blog.trailofbits.com/2025/08/07/aixcc-finals-tale-of-the-tape/)
+- [5] [CTF Radiooo AIxCC finalist interviews](https://www.youtube.com/@ctfradiooo)
+
 {{#include ../banners/hacktricks-training.md}}
diff --git a/src/AI/AI-Burp-MCP.md b/src/AI/AI-Burp-MCP.md
@@ -4,7 +4,7 @@
 
 ## Overview
 
-Burp's **MCP Server** extension can expose intercepted HTTP(S) traffic to MCP-capable LLM clients so they can **reason over real requests/responses** for passive vulnerability discovery and report drafting. The intent is evidence-driven review (no fuzzing or blind scanning), keeping Burp as the source of truth.
+Burp's **MCP Server** extension can expose intercepted HTTP(S) traffic to MCP-capable LLM clients so they can **reason over real requests/responses** for passive vulnerability discovery and report drafting. The intent is evidence-driven review (no fuzzing or blind scanning), keeping Burp as the source of truth.<sup>[[1]](#references)</sup>
 
 ## Architecture
 
@@ -17,15 +17,15 @@ Burp's **MCP Server** extension can expose intercepted HTTP(S) traffic to MCP-ca
 
 ### 1) Install Burp MCP Server
 
-Install **MCP Server** from the Burp BApp Store and verify it is listening on `127.0.0.1:9876`.
+Install **MCP Server** from the Burp BApp Store and verify it is listening on `127.0.0.1:9876`.<sup>[[3]](#references)</sup>
 
 ### 2) Extract the proxy JAR
 
 In the MCP Server tab, click **Extract server proxy jar** and save `mcp-proxy.jar`.
 
 ### 3) Configure an MCP client (Codex example)
 
-Point the client to the proxy JAR and Burp's SSE endpoint:
+Point the client to the proxy JAR and Burp's SSE endpoint:<sup>[[1]](#references)</sup>
 
 ```toml
 # ~/.codex/config.toml
@@ -43,7 +43,7 @@ codex
 
 ### 4) Fix strict Origin/header validation with Caddy (if needed)
 
-If the MCP handshake fails due to strict `Origin` checks or extra headers, use a local reverse proxy to normalize headers (this matches the workaround for the Burp MCP strict validation issue).
+If the MCP handshake fails due to strict `Origin` checks or extra headers, use a local reverse proxy to normalize headers (this matches the workaround for the Burp MCP strict validation issue).<sup>[[1]](#references)[[4]](#references)</sup>
 
 ```bash
 brew install caddy
@@ -81,7 +81,7 @@ codex
 
 ### Gemini CLI
 
-The **burp-mcp-agents** repo provides launcher helpers:
+The **burp-mcp-agents** repo provides launcher helpers:<sup>[[2]](#references)</sup>
 
 ```bash
 source /path/to/burp-mcp-agents/gemini-cli/burpgemini.sh
@@ -105,7 +105,7 @@ Example local models and approximate VRAM needs:
 
 ## Prompt pack for passive review
 
-The **burp-mcp-agents** repo includes prompt templates for evidence-driven analysis of Burp traffic:
+The **burp-mcp-agents** repo includes prompt templates for evidence-driven analysis of Burp traffic:<sup>[[2]](#references)</sup>
 
 - `passive_hunter.md`: broad passive vulnerability surfacing.
 - `idor_hunter.md`: IDOR/BOLA/object/tenant drift and auth mismatches.
@@ -133,7 +133,7 @@ Replace: User-Agent: $1 BugBounty-Username
 
 ## Burp AI Agent (AI-assisted triage + MCP tools)
 
-**Burp AI Agent** is a Burp extension that couples local/cloud LLMs with passive/active analysis (62 vulnerability classes) and exposes 53+ MCP tools so external MCP clients can orchestrate Burp. Highlights:
+**Burp AI Agent** is a Burp extension that couples local/cloud LLMs with passive/active analysis (62 vulnerability classes) and exposes 53+ MCP tools so external MCP clients can orchestrate Burp. Highlights:<sup>[[5]](#references)</sup>
 
 - **Context-menu triage**: capture traffic via Proxy, open **Proxy > HTTP History**, right-click a request → **Extensions > Burp AI Agent > Analyze this request** to spawn an AI chat bound to that request/response.
 - **Backends** (selectable per profile):
@@ -153,14 +153,14 @@ JAVA_HOME=/path/to/jdk-21 ./gradlew clean shadowJar
 # load build/libs/Burp-AI-Agent-<version>.jar via Burp Extensions > Add (Java)
 ```
 
-Operational cautions: cloud backends may exfiltrate session cookies/PII unless privacy mode is enforced; MCP exposure grants remote orchestration of Burp so restrict access to trusted agents and monitor the integrity-hashed audit log.
+Operational cautions: cloud backends may exfiltrate session cookies/PII unless privacy mode is enforced; MCP exposure grants remote orchestration of Burp so restrict access to trusted agents and monitor the integrity-hashed audit log.<sup>[[5]](#references)</sup>
 
 ## References
 
-- [Burp MCP + Codex CLI integration and Caddy handshake fix](https://pentestbook.six2dez.com/others/burp)
-- [Burp MCP Agents (workflows, launchers, prompt pack)](https://github.com/six2dez/burp-mcp-agents)
-- [Burp MCP Server BApp](https://portswigger.net/bappstore/9952290f04ed4f628e624d0aa9dccebc)
-- [PortSwigger MCP server strict Origin/header validation issue](https://github.com/PortSwigger/mcp-server/issues/34)
-- [Burp AI Agent](https://github.com/six2dez/burp-ai-agent)
+- [1] [Burp MCP + Codex CLI integration and Caddy handshake fix](https://pentestbook.six2dez.com/others/burp)
+- [2] [Burp MCP Agents (workflows, launchers, prompt pack)](https://github.com/six2dez/burp-mcp-agents)
+- [3] [Burp MCP Server BApp](https://portswigger.net/bappstore/9952290f04ed4f628e624d0aa9dccebc)
+- [4] [PortSwigger MCP server strict Origin/header validation issue](https://github.com/PortSwigger/mcp-server/issues/34)
+- [5] [Burp AI Agent](https://github.com/six2dez/burp-ai-agent)
 
 {{#include ../banners/hacktricks-training.md}}
diff --git a/src/AI/AI-llm-architecture/1.-tokenizing.md b/src/AI/AI-llm-architecture/1.-tokenizing.md
@@ -4,7 +4,7 @@
 
 ## Tokenizing
 
-**Tokenizing** is the process of breaking down data, such as text, into smaller, manageable pieces called _tokens_. Each token is then assigned a unique numerical identifier (ID). This is a fundamental step in preparing text for processing by machine learning models, especially in natural language processing (NLP).
+**Tokenizing** is the process of breaking down data, such as text, into smaller, manageable pieces called _tokens_. Each token is then assigned a unique numerical identifier (ID). This is a fundamental step in preparing text for processing by machine learning models, especially in natural language processing (NLP).<sup>[[1]](#references)</sup>
 
 > [!TIP]
 > The goal of this initial phase is very simple: **Divide the input in tokens (ids) in some way that makes sense**.
@@ -73,7 +73,7 @@ While the basic tokenizer works well for simple texts, it has limitations, espec
 
 ## Code Example
 
-Let's understand this better from a code example from [https://github.com/rasbt/LLMs-from-scratch/blob/main/ch02/01_main-chapter-code/ch02.ipynb](https://github.com/rasbt/LLMs-from-scratch/blob/main/ch02/01_main-chapter-code/ch02.ipynb):
+Let's understand this better from a code example from [https://github.com/rasbt/LLMs-from-scratch/blob/main/ch02/01_main-chapter-code/ch02.ipynb](https://github.com/rasbt/LLMs-from-scratch/blob/main/ch02/01_main-chapter-code/ch02.ipynb):<sup>[[2]](#references)</sup>
 
 ```python
 # Download a text to pre-train the model
@@ -96,7 +96,8 @@ print(token_ids[:50])
 
 ## References
 
-- [https://www.manning.com/books/build-a-large-language-model-from-scratch](https://www.manning.com/books/build-a-large-language-model-from-scratch)
+- [1] [Build a Large Language Model (From Scratch)](https://www.manning.com/books/build-a-large-language-model-from-scratch)
+- [2] [LLMs-from-scratch - ch02 tokenizer notebook](https://github.com/rasbt/LLMs-from-scratch/blob/main/ch02/01_main-chapter-code/ch02.ipynb)
 
 
 {{#include ../../banners/hacktricks-training.md}}
diff --git a/src/AI/AI-llm-architecture/2.-data-sampling.md b/src/AI/AI-llm-architecture/2.-data-sampling.md
@@ -103,7 +103,7 @@ Using the same tokenized text and `max_length` of 4:
 
 ## Code Example
 
-Let's understand this better from a code example from [https://github.com/rasbt/LLMs-from-scratch/blob/main/ch02/01_main-chapter-code/ch02.ipynb](https://github.com/rasbt/LLMs-from-scratch/blob/main/ch02/01_main-chapter-code/ch02.ipynb):
+Let's understand this better from a code example from [https://github.com/rasbt/LLMs-from-scratch/blob/main/ch02/01_main-chapter-code/ch02.ipynb](https://github.com/rasbt/LLMs-from-scratch/blob/main/ch02/01_main-chapter-code/ch02.ipynb):<sup>[[1]](#references)</sup>
 
 ```python
 # Download the text to pre-train the LLM
@@ -248,7 +248,7 @@ p(i) = \frac{w_i^{\alpha}}{\sum_j w_j^{\alpha}}
 • *w<sub>i</sub>*  – raw token percentage of corpus *i*  
 • *α* ("temperature") – a value in (0,1].  α < 1 flattens the distribution, giving more weight to smaller high-quality corpora.
 
-Llama 2 used α = 0.7 and showed that decreasing α boosted evaluation scores on knowledge-heavy tasks while keeping the training mix stable.  The same trick is adopted by Mistral (2023) and Claude 3.
+Llama 2 used α = 0.7 and showed that decreasing α boosted evaluation scores on knowledge-heavy tasks while keeping the training mix stable.<sup>[[2]](#references)</sup>  The same trick is adopted by Mistral (2023) and Claude 3.
 
 ```python
 from collections import Counter
@@ -277,12 +277,12 @@ Repeated passages cause memorization and provide an easy channel for data-poison
 2. Filter documents whose perplexity under a small reference model is > µ + 3σ (noisy OCR, garbled HTML).
 3. Block-list documents that contain PII or CWE keywords using regex & spaCy NER.
 
-The Llama 2 team deduplicated with 8-gram MinHash and removed ~15 % of CommonCrawl before sampling.  OpenAI’s 2024 "Deduplicate Everything" paper demonstrates ≤0.04 duplicate ratio reduces over-fitting and speeds convergence.
+The Llama 2 team deduplicated with 8-gram MinHash and removed ~15 % of CommonCrawl before sampling.<sup>[[2]](#references)</sup>  OpenAI’s 2024 "Deduplicate Everything" paper demonstrates ≤0.04 duplicate ratio reduces over-fitting and speeds convergence.
 
 ## Security & Privacy Considerations During Sampling
 
 ### Data-Poisoning / Backdoor Attacks
-Researchers showed that inserting <1 % backdoored sentences can make a model obey a hidden trigger ("PoisonGPT", 2023).  Recommended mitigations:
+Researchers showed that inserting <1 % backdoored sentences can make a model obey a hidden trigger ("PoisonGPT", 2023).<sup>[[3]](#references)</sup>  Recommended mitigations:
 
 * **Shuffled mixing** – make sure adjacent training examples originate from different sources; this dilutes gradient alignment of malicious spans.
 * **Gradient similarity scoring** – compute cosine similarity of example gradient to batch average; outliers are candidates for removal.
@@ -300,8 +300,8 @@ Practical recommendations:
 
 ## References
 
-- [Build a Large Language Model from Scratch (Manning, 2024)](https://www.manning.com/books/build-a-large-language-model-from-scratch)
-- [Llama 2: Open Foundation and Fine-Tuned Chat Models (2023)](https://arxiv.org/abs/2307.09288)
-- [PoisonGPT: Assessing Backdoor Vulnerabilities in Large Language Models (BlackHat EU 2023)](https://arxiv.org/abs/2308.12364)
+- [1] [Build a Large Language Model from Scratch (Manning, 2024)](https://www.manning.com/books/build-a-large-language-model-from-scratch)
+- [2] [Llama 2: Open Foundation and Fine-Tuned Chat Models (2023)](https://arxiv.org/abs/2307.09288)
+- [3] [PoisonGPT: Assessing Backdoor Vulnerabilities in Large Language Models (BlackHat EU 2023)](https://arxiv.org/abs/2308.12364)
 
 {{#include ../../banners/hacktricks-training.md}}
diff --git a/src/AI/AI-llm-architecture/3.-token-embeddings.md b/src/AI/AI-llm-architecture/3.-token-embeddings.md
@@ -185,7 +185,7 @@ For how token and positional embeddings are combined inside the model, see [the
 
 Recent work shows that context length is often limited by the positional encoding scheme rather than the token embedding matrix itself.
 
-- **Position Interpolation (PI):** Rescales position indices so longer sequences map into the range seen during training, enabling extension with minimal fine-tuning. Example:
+- **Position Interpolation (PI):** Rescales position indices so longer sequences map into the range seen during training, enabling extension with minimal fine-tuning. Example:<sup>[[2]](#references)</sup>
 
 ```python
 # Position Interpolation (PI) intuition
@@ -194,11 +194,11 @@ new_ctx = 8192
 scaled_pos = pos * (orig_ctx / new_ctx)
 ```
 
-- **YaRN:** A compute-efficient RoPE extension strategy that modifies RoPE scaling/interpolation to extrapolate to longer contexts with fewer additional training steps.
+- **YaRN:** A compute-efficient RoPE extension strategy that modifies RoPE scaling/interpolation to extrapolate to longer contexts with fewer additional training steps.<sup>[[3]](#references)</sup>
 
 ## Code Example
 
-Following with the code example from [https://github.com/rasbt/LLMs-from-scratch/blob/main/ch02/01_main-chapter-code/ch02.ipynb](https://github.com/rasbt/LLMs-from-scratch/blob/main/ch02/01_main-chapter-code/ch02.ipynb):
+Following with the code example from [https://github.com/rasbt/LLMs-from-scratch/blob/main/ch02/01_main-chapter-code/ch02.ipynb](https://github.com/rasbt/LLMs-from-scratch/blob/main/ch02/01_main-chapter-code/ch02.ipynb):<sup>[[1]](#references)</sup>
 
 ```python
 # Use previous code...
@@ -239,9 +239,8 @@ print(input_embeddings.shape) # torch.Size([8, 4, 256])
 
 ## References
 
-- [https://www.manning.com/books/build-a-large-language-model-from-scratch](https://www.manning.com/books/build-a-large-language-model-from-scratch)
+- [1] [Build a Large Language Model from Scratch (Manning)](https://www.manning.com/books/build-a-large-language-model-from-scratch)
+- [2] [Extending Context Window of Large Language Models via Positional Interpolation](https://arxiv.org/abs/2306.15595)
+- [3] [YaRN: Efficient Context Window Extension of Large Language Models](https://arxiv.org/abs/2309.00071)
 
-
-- [https://arxiv.org/abs/2306.15595](https://arxiv.org/abs/2306.15595)
-- [https://arxiv.org/abs/2309.00071](https://arxiv.org/abs/2309.00071)
 {{#include ../../banners/hacktricks-training.md}}