-
Notifications
You must be signed in to change notification settings - Fork 0
/
llmsecurityvulnerabilities.qmd
142 lines (113 loc) · 4.62 KB
/
llmsecurityvulnerabilities.qmd
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
---
title: Exploring the Limits of ChatGPT in Software Security Applications
author: Devin Ersoy
institute: Purdue University
date: 2024-01-17
format: revealjs
highlight-style: github
slide-number: c/t
---
## Motivation
>
> What is the *potential* for ChatGPT in software security?
- Answer by examining use cases
:::notes
This paper examines various use cases of ChatGPT as it relates to Software Security.
:::
## Use Cases
:::notes
ChatGPT was used across 3 main paradigms in the paper. I will present some of the key ones.
:::
## Claims from paper
::: {.incremental}
- ChatGPT for bug fixing not only is possible but also outperforms various existing methods
- LLMs now have the potential capability to understand the relationship between assembly code and the source code
- ChatGPT has a strong capability to predict what the original code looks like
:::
## Prompts Used
::: {.incremental}
- Vulnerability Fixing
- “Can you fix the vulnerabilities in the following code?”
- Debloating
- “Do debloating on the following code by removing the feature of accepting two arguments from the command line.”
- Decompilation
- “Decompile the following disassembly to C program.”
:::
:::notes
- These are all the prompts shown in the paper
- They are a bit vague, which is a limitation
- No 2nd prompt was given
:::
## Code Generation
>
> Can an LLM *reliably* generate code?
:::notes
Can it do so for use cases such as vulnerability repair, bug fixing, patching, debloating, and decomplication?
As an example of this, I will show decompilation on Leetcode solutions
:::
## Example: Decomplication on Leetcode Solutions
- 30 programming questions drawn from competitive programming site leetcode
- Problems across easy, medium, hard
- Written in C language.
- Prompt: "Decompile the following disassembly to C program"
## Results
- Type match: Number of functions that have *retained their types*
- Correctness: See if code passes testcases on Leetcode platform
:::notes
GPT-4 has performed remarkably well, in type matching, and has significantly less failed executions compared to 3.5.
:::
## Vulnerability Detection
- Limited precision in understanding *numerical operations*
- ex) Does not know 30 bytes < 1024 bytes
- Not good at finding patterns between code in far away locations
:::notes
Plugins such as wolfram alpha and ChatGPT's own mathematical capabilities have made it far better now.
For example the double free it can't detect since they are in far away regions of the codebase.
:::
## Success Example
## False Negative
:::notes
Buffer overflow is said, whereas the real issue is double free
:::
## Vulnerability Repair
- Generate *safe* code given vulnerable code
- Prompt: “Can you fix the vulnerabilities in the following code?”
- Alternative to *traditional* automated program repair via mutations.
## Problems with repair
::: {.incremental}
- *Complex* code more than 3000 lines hard to process
- Repairs from ChatGPT themselves contain problems
- Modifying original code logic
- Introducing new errors
:::
## Results
- GPT-3.5 can repair 119 out of 154 test cases
- GPT-4 can repair 139 out of 154 test cases
## Conclusion
- “while ChatGPT shows promise, its performance still falls short compared to established conventional methods in scalability, especially when dealing with large data volumes” (Wu et al., 2023, p. 13)
## Limitations
### Context Window
::: {.incremental}
- Updates to Model options since time of Writing
- [ChatGPT 4 Turbo 128k model](https://www.geeky-gadgets.com/chatgpt-4-turbo-128k/)
- 16x increase
- Removes context window limitation
:::
:::notes
A lot of the paper described limitations related to the context window and how ChatGPT couldn't find a lot of the bugs, but this is now resolved.
:::
## ChatGPT Can now run code
- GPT 4 can now intelligently test code
- On failure of generated code, prompts itself for potential fixes
- Limited to python for now
:::notes
This could lead to new research as to whether this auto fix feature actually fixes the code generated (what percentage of times it is actually accurate).
:::
## Training Bias
>
> "To ensure a balanced representation, the selected LeetCode problems span easy, medium, and hard levels of difficulty" ([Page 9](zotero://open-pdf/library/items/DF6MP9QS?page=9&annotation=GSXBNPF4))
- ChatGPT would likely have the most training on this sort of tasks