Understanding the fundamentals of attacks

Understanding the fundamentals of attacks
What is happening when someone writes an exploit?
Halvar Flake / Thomas Dullien
November 3, 2016
The /home owners association

Table of contents
1. The need to deﬁne “exploit”
2. Deﬁning bug-free and buggy software
3. Intended states, CPU states, and weird states
4. Input streams as instruction streams
5. Provably exploitable and provably unexploitable code
6. What does all of this mean?
7. Summary
1

The need to deﬁne “exploit”

What, exactly, is an ‘exploit’?
What, exactly, is an exploit?
Central concept in computer security.
Valuable currency: Used to be barter-traded, now is traded for $$$.
Could you deﬁne it so that an old-school computer scientist understands?
2

What, exactly, is an ‘exploit’?
In spite of centrality of the concept, it is ill-deﬁned.
Informal deﬁnitions abound, from overly abstract to overly concrete:
“A program that allows you to do something that you are not
supposed to be able to do.”
“I recognize an exploit when I see it.”
“Magic”
“When you overwrite EIP”
3

Why is the lack of a definition a problem?
Problem 1:
Since we do not have clear definitions, we fall into scholastic arguments:
“Adding a cookie to the heap here will stop exploits - it breaks
the technique described in Phrack!”
“It is just a single bit error, and hence not exploitable.”
Since we do not have clear definitions, these statements cannot be
reasoned about properly.
4

Computer security is full of non-scientiﬁc statements, assertions, and
unquantiﬁable handwaving.
Mitigations are supposed to “raise the bar”, but:
Bar-raising
1. How much will the bar be raised?
2. Will it cost the attacker time once, or will it cost him time on each
exploit?
3. Is the cost (performance, complexity) worth the gain?
Our discussions are similar to “how many angels can balance on the tip
of a needle”.
5

Most defenders have no first-hand experience with a modern exploit.
With security jobs doubling every 4 years, scholasticism is a real problem.
Incentive structures are harmful:
Happy defender
1. Come up with overspecific
countermeasures
2. Show progress to superiors
& get promoted
Happy attacker
1. Bypass overspecific
countermeasures
2. Hack all the things
3. Show progress to superiors
& get promoted
6

Problem 2:
We fail at teaching it properly, and instead teach a “bag of tricks”
The broken exploitation curriculum:
1. Learn about stack smashing
2. Learn about heap implementation X, Y, and Z
3. Learn about SQL injection
4. Learn about ASLR and other mitigations
5. Learn about XSS
6. Learn about other overly special cases
Disjoint, disparate, and terrible at transmitting principles. Too many
people “miss the forest for the trees” as a result.
7

Why is there no good theoretical body on exploits?
“Philosophy of science is about as useful to scientists as
ornithology is to birds” (Feynman)
Exploit practitioners know the intuitions they need, and have no incentive
to formalize or disseminate their mental models.
8

This talk
Explains the “right” way to think about exploits and exploit writing1
.
Introduces an attacker model for memory corruptions that allows us to
reason about large classes of attacks.
Discusses a few theoretical results & their consequences.
Omits proofs and extreme formalism.
... but remains a theoretical CS talk, so don’t fall asleep!
1“Right” as subjectively judged by me, but I have impeccable taste in exploits.
9

Credit where it is due
Intuitions behind this talk are known to many exploit practitioners.
“Folk theorems” - results are known, but not formally written down
anywhere.
The LANGSEC2
community have contributed strongly to the discussion
and evolution of ideas.
2L. Sassaman, M. L. Patterson, S. Bratus, J. Vanegue etc. ...
10

Deﬁning bug-free and buggy
software

Recap: Finite state machines
A graph of states, and input drives the machine along edges
(“transitions”) to other states.
State 1 State 2 State 3
‘A’ ‘B’
‘A’
‘D’ ‘E’
If you are unfamiliar with ﬁnite-state machines, think of such diagrams: 11

The intended (finite state) machine
Why write software? Because you wish to address a problem.
You want to have a machine that does something specific for you.
Ideally, you want to have a specific finite-state machine that addresses
your problem.
12

Unfortunately, you do not have this machine: All you have is a
general-purpose computer.
So you build an emulator that emulates the machine you would like to
have on a general-purpose computer.
Programming is, at its heart, construction of ﬁnite-state-machine
emulators. 3
3Some people may disagree, but it is true for ﬁnite-RAM (e.g. real) machines
13

The machine that addresses the problem is the intended finite state
machine: The machine you wanted to build.
By definition, it is bug-free: It does exactly what you want it to do, and
nothing else.
All theory is grey, so let’s define a small IFSM: A system that stores 5000
password / secret-value combinations
14

The IFSM
A simple secret-keeping machine:
A: Read password/secret
B: Store pair in memory
C:
Remove pair from memory
Output the requested secret
D: Output error message
New password?
Fewer than 5000 stored?
Pass/secret not equal 0?
Existing password?
Pass/secret equal 0?
Already 5000 stored?
15

The IFSM, slightly more formal
A simple secret-keeping machine:
Read input password/secret
read(p)
read(s)
Store pair in memory (B)
Memory ← Memory ∪ {(p, s)}
Output the requested secret (C)
Memory ← {(p , s ) ∈ Memory | p = p}
print(s )
Output error message (D)
print(0)
IF condition b:
∀(p , s ) ∈ Memory : p = p
|Memory| ≤ 4999
s = 0, p = 0
IF condition c:
∃(p , s ) ∈ Memory : p = p
s = 0
IF condition d:
s = 0
∨p = 0
∨ |Memory| = 5000
16

Implementing the IFSM
This machine can be implemented in many ways.
We will examine 2 diﬀerent implementations.
Memory in the IFSM is a set. Computers do not know sets.
Our two implementations will emulate Memory via (1) a ﬂat array and
via (2) a linked list.
17

Flat-array implementation: Writing
Divides memory into pairs of memory cells.
Pairs are either empty (and both 0), or represent secret/password combo.
Insertion is linear scan for empty pair and writing there.
Removal is linear scan for correct pair and removing it.
We refer to such a pair as (p, s).
18

Transition (B): Memory ← Memory ∪ {(p, s)}.
Search for ﬁrst pair with p = 0, write data there.
19

20

P S
21

P S
22

Flat-array implementation: Reading
Transition C: Memory ← {(p , s ) ∈ Memory|p = p}.
Searching for a speciﬁc p and removing the pair.
P S
23

P S
Linear sweep until right p is found.
24

P S
Linearly sweep until right p is found. Output corresponding s.
25

Linearly sweep until right p is found. Output corresponding s. Zero
memory cell.
26

Linearly sweep until right p is found. Output corresponding s. Zero
memory cell.
27

The linked-list implementation
Linearly sweeping memory is computationally expensive and not very
CSy.
Let us replace the ﬂat array with a linked list for free and used tuples.
Divide memory into 3-tuples: (p, s, next). Keep track of two list heads
for free and used tuples.
28

Linked-List implementation
Keep all used blocks in a linked list.
30

Keep all free blocks in a linked list.
31

Keep two list heads in registers.
32

Linked-list implementation details
Very detailed step-by-step slides for what happens are in the appendix,
but most people will be familiar with linked-list unlinks.
We will look at the machine implementation in more detail later.
Right now, we need to deﬁne what security properties are.
33

Security properties of the IFSM
The security properties of the IFSM are “what we want to be true” for
the IFSM.
This is needed to deﬁne “winning” for an attacker: He wins when he
defeats the security properties of the IFSM.
The security property is just a statement about input and output
sequences of the IFSM.
34

Security properties of the IFSM
A secret-keeping machine should not allow the secrets to be retrieved by
someone who does not know the password.
“If the system stores a pair (p, s) it should be practically infeasible for
the attacker to obtain s if he does not know p”.
35

Intended states, CPU states, and
weird states

Intended states
There is one IFSM state for each possible conﬁguration of Memory.
At the same time, we saw that there can be multiple states of the CPU
that represent the same IFSM state (sets do not care about ordering).
You can map between intended states and sets of CPU states: Every
state of the IFSM can be represented by one or more CPU states.
But many CPU states do not represent any state of the IFSM.
36

CPU states
CPU states can fall into one of three categories:
Sane states: A CPU state which corresponds to a state of the IFSM.
Transitory states: A CPU state which inevitably irrespective of further
input or output leads to a sane state. These occur when the emulator
emulates a transition arrow of the IFSM using multiple instructions.
Weird states: A CPU state that does not correspond to a state of the
IFSM, and hence cannot be interpreted as a representation of the IFSM.
37

Weird states
How can the CPU enter a weird state?
Programming mistakes
Bitﬂips in DRAM
Broken data on your harddisk that gets paged in
... and a myriad of other ways ...
38

Emulated transitions on weird states
The program that executes on the CPU usually cannot know it has
entered a weird state.
It will happily continue executing until an exception occurrs.
This means that it will emulate transitions intended for a sane state on a
weird state.
This transforms one weird state into (usually) another weird state.
39

Emulated transitions on weird states
Once a weird state is entered, many other weird states can be reached by
applying the transitions intended for sane states on them.
A new computational device emerges, the weird machine.
The weird machine is the computing device that arises from the
operation of the emulated transitions of the IFSM on weird states.
...time for a radical change of perspective...
40

Input streams as instruction
streams

Normal model of computation
State 1 State 2 State 3 State 4 State 5
Input Input
Instruction Instruction Instruction Instruction
Program
Data
41

Normal model of computation
Summarize the middle transitions.
Input Input
Instruction(s) Instruction(s)
Program
Data
42

Attacker model of computation
Input Input
Instruction(s) Instruction(s)
Data
Program
43

Switching perspective
Any computing device that accepts input and reacts to this input by
executing a diﬀerent program path can be viewed as a computing device
where the inputs are the program.
If no weird state can ever be entered, this is no problem, and intended.
If there is a way to enter a weird state, all hell breaks loose: The attacker
is now programming your computer.
44

The essence of exploitation
Given a method to enter a weird state from a set of particular sane
states, exploitation is the process of
setup (choosing the right sane state),
instantiation (entering the weird state) and
programming of the weird machine
so that security properties of the IFSM are violated.
45

Provably exploitable and
provably unexploitable code

Attacker model
Cryptography has a variety of attacker models of diﬀerent strengths.
Plaintext-only. Chosen-ciphertext. Chosen-plaintext. Adaptive
chosen-plaintext.
These attacker models deﬁne what an attacker is allowed to do and
generalize many real-world attacks.
They are usually quite powerful, and allow reasoning about large classes
of attacks and defenses.
46

Attacker model
Let us introduce a simple (yet powerful) attacker model for memory
corruptions:
A chosen-bitﬂip-once attacker is allowed to corrupt exactly one bit of
memory exactly once, when the machine is waiting for input.
There are many other attacker models that are sensible and imaginable -
but we won’t discuss them here.
47

Attacker model
Our attacker model works as follows:
1. Attacker gets to send inputs to the machine
2. Defender gets to send an input (p, s) to the machine
3. Attacker gets to send inputs to the machine
4. Attacker gets to ﬂip exactly one bit of his choosing
5. Attacker gets to send more inputs to the machine to obtain s
The chosen-bitﬂip-once attacker is a fairly strong attacker model, but not
unreasonably so.
48

Proving exploitability of the linked-list variant
The time-honored way of showing exploitability is providing an exploit.
We follow the same route: We describe a sequence of operations that
allows the attacker to obtain the secret.
49

Exploiting the linked-list implementation
Initial state is a list of empty 3-tuples.
free
head
=0
50

Attacker sends (p0, s0), (p1, s1), (p2, s2) sequence.
free
head
=9
p0 s0 n0 p1 s1 n1 p2 s2 n2
51

Attacker sends (p0, s0), (p1, s1), (p2, s2) sequence. Defender sends
(pd , sd ).
p0 s0 n0 p1 s1 n1 p2 s2 n2 pd sd nd
free
head
=12
52

(pd , sd ). Attacker sends (p2, X).
free
head
=6
53

(pd , sd ). Attacker sends (p2, X), (p1, X).
free
head
=3
54

(pd , sd ). Attacker sends (p2, X), (p1, X), (p3, s3).
free
head
=6
55

(pd , sd ). Attacker sends (p2, X), (p1, X), (p3, s3), (p4, s4).
free
head
=12
56

The machine is still in an entirely sane state.
free
head
=12
57

The attacker gets to corrupt one bit, incrementing n3.
free
head
=12
58

The attacker gets to corrupt one bit, incrementing n3. He then sends in
(s4, X).
free
head
=12
59

The machine follows the linked list, interprets the stored s4 as password,
and outputs n4.
free
head
=12
60

and outputs n4. It then sets the three cells to be free and overwrites the
stored pd with the free-head.
p0 s0 n0 p3 s3 n3 p4 s4 n4 12 sd nd
free
head
=7 61

and outputs n4. It then sets the three cells to be free and overwrites the
stored pd with the free-head. Attacker sends (12, X) and obtains sd .
p0 s0 n0 p3 s3 n3 p4 s4 n4 12 sd nd
free
head
=7 62

Proving non-exploitability of the flat-array variant
The proof-of-exploitability was constructive.
Proving non-exploitability has a different flavor.
It is simply enumerating all initially reachable weird states and showing
they lead nowhere.
63

By flipping a bit, an attacker can enter one of the following states:
1. A sane state where he modifies a previously-saved p by one bit.
2. A sane state where he modifies a previously-saved s by one bit.
3. A weird state where a password-field of a (0, 0)-pair is set to = 0
but the secret field of the pair is 0.
4. A weird state where a secret-field of a (0, 0)-pair is set to = 0.
The sane states do not give him any advantage in obtaining sd . What
about the weird states?
64

Examine the initial weird states:
3. A weird state where a password-field of a (0, 0)-pair is set to = 0
but the secret field of the pair is 0.
3.1 The machine can be made to spit out 0 given the right password,
then reverts to a sane state.
3.2 This is weird behavior, but does not violate security properties
4. A weird state where a secret-field of a (0, 0)-pair is set to = 0.
4.1 The machine will treat the pair as empty and overwrite it, reverting
to a sane state.
4.2 This is weird behavior, but does not violate security properties
The attacker is out of options.
65

Interpreting all of this (1 of N)
Exploitation has very little to do with hijacking RIP/EIP.
All programs we considered had perfect control-ﬂow integrity, and no
instruction pointer was hijacked.
Exploitation is programming a weird machine.
66

Properties of weird machines
Weird machines get more powerful as the emulator for the IFSM gets
more complex.
It seems that the presence of a linked list in the same address space
where a bit gets ﬂipped may already imply “anything can happen” if the
attacker can cause unlink/link operations.
Weird machines are created out of the IFSM emulator. They are not
designed to be programmed, and the attacker writes “weird machine
assembler”.
67

Properties of weird machines
Since the weird machine arises from its “host”, diﬀerent versions of the
same “host” may create similar weird machines.
This leads to attacker specialization: “X is the expert for IE exploitation.”
Attackers also end up writing libraries for recurrent exploitation of the
same targets. Nobody wants to write assembler from scratch, especially
not weird machine assembler.
68

Many mitigations break one particular weird machine program.
Many mitigations can be “programmed around” by the attacker.
Many mitigations can be “bypassed once” (for a given target), then the
attacker has a “library” to repeat the feat when he ﬁnds the next bug.
69

It is actually possible to write non-exploitable programs.
Exploring the space of non-exploitable programs seems worthwhile.
What datastructures can be implemented in non-exploitable manner?
At what computational cost?
70

“Provably correct” is not the same as non-exploitable. These are
orthogonal concepts.
“Provably correct” means there is no software-only path to induce a
weird state.
Non-exploitable means all weird states a given attacker can reach resolve
to harmless sane states.
71

“Turing-complete” is a fashionable way of showing the power of an
oﬀensive technique in some academic papers.
It is a misuse of the term. It is trivial to produce a IFSM that simulates a
two-counter machine that is “Turing-complete”, but no security
properties are violated.
The right way of showing power is demonstrating violation of security
properties.
If you want to show more, demonstrate violation of all possible security
properties. So you want to show you can reach any pathological state.
(Turing-complete computation on arbitrary memory can do this).
72

When trying to write non-exploitable programs, it is very hard to keep
“indirection” under control.
If one memory cell refers to another (directly or indirectly), it is often not
easy to assure it remains sane.
Theoretical computer science distinguishes between RAM machines with
indirect addressing and without. Is something similar at play here?
73

Summary
Exploitation is programming emergent weird machines.
It does not require EIP/RIP, and is not a bag of tricks.
Theory of exploitation is still in embryonic stage.
A lot of interesting questions - where will we stand in 10 years?
74

Acknowledgements
A lot of people have discussed the topic with me and thus inﬂuenced my
views. In random order: Felix Lindner, Ralf-Philipp Weinmann, Willem
Pinckaers, Vincenzo Iozzo, Julien Vanegue, Sergey Bratus, William
Whistler, Sean Heelan, Sebastian Krahmer.
A lot of thanks for their patience and insight.
I have spoken to so many people about exploitation in recent years. If
you feel you should’ve been listed here, just assume you were :-)
I also have to thank Mara Tam for useful feedback and outside
perspective.
75

References
Creating a proper full bibliography is diﬃcult and out of scope for this
conference talk. Some starting points:
[3] [2] [1]

References I
S. Bratus, J. Bangert, A. Gabrovsky, A. Shubina, M. E. Locasto, and
D. Bilar.
’weird machine’ patterns.
In C. Blackwell and H. Zhu, editors, Cyberpatterns, Unifying Design
Patterns with Security and Attack Patterns, pages 157–171.
Springer, 2014.
L. Sassaman, M. L. Patterson, S. Bratus, and M. E. Locasto.
Security applications of formal language theory.
IEEE Systems Journal, 7(3):489–500, 2013.
J. Vanegue.
The weird machines in proof-carrying code.
In IEEE Symposium on Security and Privacy Workshops, pages
209–213, 2014.

Linked-list implementation
Take ﬁrst element from free-list head.

Linked-list implementation: Writing
How do we store something in memory?
Take ﬁrst element from free head. Adjust free-list head.

Take ﬁrst element from free head. Adjust free-list head. Fill in p, s.

Take ﬁrst element from free-list head. Adjust free-list head. Fill in p, s.
Fill in next from the used-list head.

Fill in next from the used-list head. Adjust used-list head.
Fill in next from the used-list head. Adjust used-list head.

Fill in next from the used-list head. Adjust used-list head. Done.

Meditative slide for everybody to rest for a minute.

Follow used-list until p is found.

Follow used-list until p is found. Look at next.

Follow used-list until p is found. Adjust next that led us here so that the
block with p is removed from the used-list.

Output s, and replace next with free-list head.

Output s, and replace next with free-list head. Adjust free-list head.

Understanding the fundamentals of attacks

Recommended

More Related Content

What's hot (20)

Viewers also liked (20)

Similar to Understanding the fundamentals of attacks (20)

More from Cyber Security Alliance (20)

Recently uploaded (20)

Understanding the fundamentals of attacks