BLASTPASS SANS HackFest
Document
The BLASTPASS SANS HackFest presentation focuses on a zero-click zero-day exploit chain that affects iOS and macOS devices, which involves a series of vulnerabilities including a flaw in PassKit, the blastdoor security feature, Address Space Layout Randomization (ASLR), Pointer Authentication Code (PAC), and a buffer overflow vulnerability in the Huffman tables used in WebP compression.
Full Transcript / Content
--- Page 1 ---
--- Page 2 ---
BLASTPASS
CVE-2023-41061 & 41064
Overview & Exploit Chain
Threat Briefing
SOC-Class.com || Montance® LLC 2Copyright 2023 Montance® LLC - All Rights Reserved. All Wrongs Reversed?
--- Page 3 ---
Christopher Crowley – Montance® LLC
- Background: root on
(Ultrix/VMS) employer systems
at 15 years old (80s #CYBER)
- Sectors: Defense, Education,
Energy, Government, Software
Development, Financial,
Telecom…
- SANS Senior Instructor.
Consultant, SOC-Class.com
author. Teaches: SEC575,
SEC511, SEC504,…
Montance® LLC
3
--- Page 4 ---
Christopher Crowley – Montance® LLC
- Slide deck download:
https://montance.com/pres
- SANS website will also have it
- Please use this information for
yourself, your organization, and
to help your community
- @CCrowMontance (infrequent)
- https://mgt517.com/linkedin
Montance® LLC
4
--- Page 5 ---
Quick History:
CVE-2021-30860 - FORCEDENTRY
SOC-Class.com || Montance® LLC 5Copyright 2023 Montance® LLC - All Rights Reserved. All Wrongs Reversed?
--- Page 6 ---
Information Sources
- There’s a history of Citizen Labs tracking NSO:
- Citizen Labs FORCEDENTRY 2021-09-13 blog:
https://mgt517.com/cl-forcedentry
earlier 2021-08-24 blog post:
https://mgt517.com/cl-forcedentry-0
- Apple Security Update 2021-09-13:
https://mgt517.com/apple-forcedentry
^^^ Link shortener links are case sensitive
--- Page 7 ---
Zero-click Zero Day
- Dangerous zero-click zero-day exploit discovered
- Zero-day: exploit found without a patch available
- Zero-click: user interaction not required
- Patch Now! for CVE-2021-30860 and CVE-2021-
30858 to defend iOS, macOS, watchOS against
the exploit
- Zero day exploit code dubbed FORCEDENTRY by
Citizen Labs (who disclosed it to Apple)
--- Page 8 ---
Attribution: NSO
Trivia: a Pegasus is not a Unicorn
SOC-Class.com || Montance® LLC 8Copyright 2023 Montance® LLC - All Rights Reserved. All Wrongs Reversed?
--- Page 9 ---
Attribution in General
- First, attribution is tricky business and may not make
much difference to you. Would your actions change if it
was NSO or NSA who developed this?
- The disclosure of a zero day means that other capable
threat actors are aware of this flaw and may quickly
develop an exploit prior to patch deployment
- Pegasus install cited in Citizen Labs post, other details
withheld in an OPSEC maneuver to aid victim discovery;
likely also substantial unknowns on the rest of the details
(** will discuss their previous data)
--- Page 10 ---
Attribution in General
- US Defense Science Board schema: $T, $M, $B
- Reuse, discover, and create/introduce flaws
- This is the work of $M level actor
- Difficult and damaging, but not an
existential threat for most,
nor the creation of a flaw
- Sad state of affairs, but zero-days
happen, you should already be ready
to mobilize and shift gear
$B
create
$M
discover
$T
reuse
--- Page 11 ---
Value of Attribution
- If you are in a position to (or have to regardless)
investigate, there are a large number of specific IOCs
posted in the Citizen Labs blogs
- Previously cited operators of Pegasus infrastructure, and
the specific IP address deployments of those operators
- Review IOCs from CL post, also @craiu (Kaspersky)
detailed thread on twitter for attachment inspection:
https://mgt517.com/forcedentry-forensic
- This list of resources is constantly shifting
--- Page 12 ---
Citizen Labs Basis for NSO Attribution
- Phone was infected with Pegasus spyware
- NSO is a software provider, Pegasus is its product
- FORCEDENTRY (’21) CASCADEFAIL:
- Incomplete deletion of data from DataUsage.sqlite
- Rationale: to reduce evidence of spyware activity, the
ZPROCESS table is modified, however the ZLIVEUSAGE table
isn’t modified and the discrepancy is useful
- Citizen Lab blog says they’ve only observed this from Pegasus;
SQL query when performing forensic analysis to determine if
the phone you are assessing is exhibiting same behavior:
SELECT "CASCADEFAIL" FROM ZLIVEUSAGE WHERE ZLIVEUSAGE.ZHASPROCESS NOT IN (SELECT Z_PK FROM ZPROCESS);
--- Page 13 ---
Citizen Labs Basis for NSO Attribution
- FORCEDENTRY (‘21) setframed:
- Distinctive process name “setframed” as part of the July
2020 Al Jazeera targeted attacks (also attributed by Citizen
Lab to NSO) forensic analysis performed. Citizen Labs
blogpost indicates this detail wasn’t publicly disclosed
previously
- Blog indicates this is a process name associated with this
intrusion spyware
--- Page 14 ---
Citizen Labs Basis for NSO Attribution
- BLASTPASS (‘23):
- “Citizen Lab found an actively exploited zero-click vulnerability
being used to deliver NSO Group’s Pegasus mercenary
spyware.”
- “We expect to publish a more detailed discussion of the
exploit chain in the future.”
From: https://mgt517.com/blastpass
- No details, technical corroboration, etc.
- I asked for more details: no response; didn’t expect it tbh
- So, confidence and trust based on previous assessments
--- Page 15 ---
BLASTPASS:
Limited Exploit Details
SOC-Class.com || Montance® LLC 15Copyright 2023 Montance® LLC - All Rights Reserved. All Wrongs Reversed?
--- Page 16 ---
Relationship between
CVE-2023-41061 &
CVE-2023-41064
SOC-Class.com || Montance® LLC 16Copyright 2023 Montance® LLC - All Rights Reserved. All Wrongs Reversed?
--- Page 17 ---
41061: Wallet / 41064: ImageIO
- Citizen labs found the exploit, did preliminary
analysis, provided details to Apple
- Citizen Lab cited as CVE-2023-41064 credit
- Apple appears to have identified the
blastdoor circumvention (CVE-2023-41061)
within PassKit
- CVE-2023-41064 is also be related to CVE-
2023-4864 (will discuss later)
--- Page 18 ---
CVE-2023-41061
Validation Issue Apple Wallet
SOC-Class.com || Montance® LLC 18Copyright 2023 Montance® LLC - All Rights Reserved. All Wrongs Reversed?
--- Page 19 ---
iOS Wallet / PassKit Update
Wallet: “A validation issue was addressed with
improved input validation.”
- CVE credit is to “Apple”
(patch announcement for iOS 16.6.1:
https://mgt517.com/41061)
--- Page 20 ---
CVE-2023-41064
Buffer Overflow ImageIO
SOC-Class.com || Montance® LLC 20Copyright 2023 Montance® LLC - All Rights Reserved. All Wrongs Reversed?
--- Page 21 ---
Buffer Overflow
- Buffer overflows: a category of programmatic
flaw. If the space allocated to store process
the result value is inadequate the resultant
write of data can result in overflowing a
buffer in memory.
- This fix is addressing the starting point of a
zero-click exploit chain, let’s look into that in
a little more detail
--- Page 22 ---
CVE-2023-41064: ImageIO
- Buffer Overflow patched by CVE-2023-41064
- Citizen lab blog cites malicious images sent
via iMessage to victim, and the corresponding
PassKit flaw as part of the exploit chain
- https://mgt517.com/blastpass
- iMessage delivery > PassKit processing
outside blastdoor > ImageIO exploit
--- Page 23 ---
Zero clicks
I give…
SOC-Class.com || Montance® LLC 23Copyright 2023 Montance® LLC - All Rights Reserved. All Wrongs Reversed?
--- Page 24 ---
Zero Clicks
- For mobile devices, there isn’t a service side
opportunity for remote code execution
- Mobile Attack surface preso I gave in 2020:
https://mgt517.com/mobile-attack
- Attackers seeking remote code execution are
largely relegated to messaging apps to
process content without user interaction
(anyone remember Stagefright?)
--- Page 25 ---
Blastdoor (1 of 3)
- In response to previous zero click exploits,
Apple developed blastdoor for iOS 14
- This is a “tainted data” processor which
severely limits the processing done on data
sent over SMS/iMessage as ingested to phone
- Based on Google Project Zero writeup
- https://mgt517.com/p0-blastdoor
- BLASTPASS name: PassKit violating blastdoor
--- Page 26 ---
Blastdoor (1b of 3)
From: https://mgt517.com/p0-blastdoor
(Speculated) PassKit
involvement required
here: cue 41061
--- Page 27 ---
Blastdoor (2 of 3)
- (‘21) apsd -> identitydservicesd -> imagent ->
MessageBlastDoorService -> imagent ->
IF:serializedData? -> IMTranscoderAgent <<<
- (‘21) In the “CL-Pearls” writeup they discuss two
types of crashes involving IMTranscoderAgent
- Type1: copyGifFromPath:toDestinationPath:error in
ImageIO to render .psd files (27 of the same sent)
- Type2: copyGifFromPath:toDestinationPath:error in
CoreGraphics for decoding JBIG2 data in PDF file
(4 different PDF files sent)
--- Page 28 ---
Blastdoor (3 of 3)
- PassKit details for pass parsing
- Little public material available
- Similar scenario as in FORCEDENTRY
- Special attachment type (wallet) requires moving data
out of blastdoor protected data segment into the wallet
context, which enabled the exploitation to take
advantage (this is speculated / unverified)
- Satisfying item #4 of the list of items (on the next slide)
--- Page 29 ---
Zero Click Necessary Components
- According to Project Zero post, zero-click needs:
- Memory corruption vulnerability (buffer overflow in ‘23)
- Remotely break ASLR (unknown currently)
- Code execution (unknown currently)
- Sandbox escape (PassKit in ‘23)
- Pointer Authentication Code (PAC) evasion
- To overcome blastdoor’s protections: overseeing the
process of ingesting iMessages, dyld shared cache
rerandomized for each service start, ASLR brute
mitigation by doubling restart times between each crash
--- Page 30 ---
Zero Click Necessary Components
- According to Project Zero post, zero-click needs:
- Sandbox escape (Presumed: PassKit in ‘23)
- My speculation on the PassKit and related sandbox
specifications is there’s some service (XPC seems
reasonable) that enables transfer of attachment to
PassKit and starts the exploit
--- Page 31 ---
Zero Click Necessary Components
Memory
corruption
vulnerability
(CVE-41064
in ‘23)
Remotely
break ASLR
(unknown
currently)
Code
execution
(unknown
currently)
Sandbox
escape (CVE-
41061 /
PassKit in
‘23)
PAC
(discussed in
p0 blog
post)
--- Page 32 ---
PAC
- That blog post indicates that pointer
authentication codes (PAC) weren’t an issue for
that specific exploit under assessment (iOS 12.4)
- ARM protection as of ARM v8.3 that associates
code calling code to protect against ROP / JOP
style execution
- Requires exploit to develop primitives in
complicated ways to assure the calls are
accepted as valid (I.Beer post on ImageIO IPC)
--- Page 33 ---
Exploit Info & Speculation
- Synthesizing info from two blog posts from
citizen labs, Apple report, the Project Zero
blogpost on the functional teardown of
Blastdoor, PassKit flaw
- We have an understanding of the majority of the
components necessary to accomplish the CitizenLab
reported
- Buffer overflow cited
- Movement from PassKit flaw to WebP
--- Page 34 ---
CVE-2023-41064
(and CVE-2023-4863)
Ready for the curveball?
SOC-Class.com || Montance® LLC 34Copyright 2023 Montance® LLC - All Rights Reserved. All Wrongs Reversed?
--- Page 35 ---
WebP Lossless Encoding Patch
- Presuming PassKit’s “validation issue” was
leveraged to get to ImageIO, scant details related
to 41064, except for something interesting…
- ImageIO supports WebP; Apple reported WebP
bug on Sept. 6, 2023; urgent Chrome patch
- Ben Hawkes, “Based on this, it seems likely that
the BLASTPASS vulnerability and CVE-2023-4863
("the WebP 0day") are the same bug.”
- https://mgt517.com/hawkes-webp
--- Page 36 ---
WebP Algorithm Logic Flaw
- WebP fix details: https://mgt517.com/webp-oob
- Huffman table (tree structure originally) used to
perform lossless compression
- Vulnerable webP code is from buffer sizes defined
based on fixed size tables for memory allocation;
resulting in overflow
- Fix checks size; allocates space
for larger requests
--- Page 37 ---
Down the WebP-Hole (1 of 3)
- Processing is intended to deliver lossless image
encoding and provide compression to reduce size
- Starts getting mathy… to reduce image size,
frequency analysis is performed on image pixels
- Specific output bits are mapped to a tree tracking
sequences, frequently occurring bits are reduced
to smaller sequences than the original bits
- The bit length provides a unique mapping, but the
Huffman table is needed to recover original data
down the rabbit hole, but make it
a webP hole instead of rabbit,
using Huffman encoding
Image from: dream.ai
--- Page 38 ---
Down the WebP-Hole (2 of 3)
- Which means the size would increase, so the
Huffman tables are also compressed the same way
- Hawkes hypothesized (VP8L is another name for
webp) VP8LBuildHuffmanTable/BuildHuffmanTable
method call would be the candidate for shifting
huffman_table pointer past pre-calculated buffer
size (since the pre-calculation was fixed in the patch
this is targeted)
- He then tried to create a table that would create an
exploit situation
down the rabbit hole, but make it
a webP hole instead of rabbit,
using Huffman encoding
Image from: dream.ai
--- Page 39 ---
Down the WebP-Hole (3 of 3)
- As it turns out, even with some related code (Mark
Adler enough.c referenced by Google) he couldn’t
create a table that was vulnerable
- enough.c was only 8-bit color code, so dead end
- Spends lots of time creating tables, but couldn’t
create invalid tables
- Modifies the ReplicateValues, finally established a
condition creating an invalid table
down the rabbit hole, but make it
a webP hole instead of rabbit,
using Huffman encoding
Image from: dream.ai
--- Page 40 ---
WebP-Hole Exploit POC
- In order to establish the invalid table, he’s
dealing with the last 5 tables, but can’t
generate tables big enough to cause errors
- So, the POC builds 4 tables to fill the space
expected for the 5 tables, then adds a final
invalid table with ReplicateValue which
causes the out of bounds write to occur .
down the rabbit hole, but make it
a webP hole instead of rabbit,
using Huffman encoding
Image from: dream.ai
--- Page 41 ---
WebP-Hole Exploit POC
- This creates the invalid fifth table, corrupting
memory, but still within the constraints of
the math for the Huffman tables
- From craft.c:
down the rabbit hole, but make it
a webP hole instead of rabbit,
using Huffman encoding
Image from: dream.ai
--- Page 42 ---
WebP-Hole Exploit POC
dwebp output:
down the rabbit hole, but make it
a webP hole instead of rabbit,
using Huffman encoding
Image from: dream.ai
Hawkes Post
@mistymntncop POC
--- Page 43 ---
WebP-Hole Exploit POC
Dwebp output:
down the rabbit hole, but make it
a webP hole instead of rabbit,
using Huffman encoding
Image from: dream.ai
Hawkes Post
@mistymntncop POC
--- Page 44 ---
WebP Pervasive Attack Surface
- WebP attack surface is extensive
- CVE-2023-4863 (chrome) and CVE-2023-5129 (later rejected
as they pushed libwebp generally back into 4863)
- At the time of writing of several of the initial analyses, it
was speculated that Android was vulnerable to it
- Acknowledged in
October update:
--- Page 45 ---
WebP Flaw Discovery Speculation
- Hawkes then talks about trying to find this bug
through fuzzing and says none of the current
fuzzers could do it; speculates manual effort
- He built the invalid table because of the patch
- A mathy-math muthamather developed the
exploit as a novel concept against Huffman tables
- My thoughts: WebP is relatively new, in ImageIO,
open source, so good target / attack surface
- https://mgt517.com/hawkes-webp
--- Page 46 ---
Unknowns & Expected
Forthcoming Information
SOC-Class.com || Montance® LLC 46Copyright 2023 Montance® LLC - All Rights Reserved. All Wrongs Reversed?
--- Page 47 ---
Speculation on Future Patches
- We know a fraction of the details
- We don’t know:
- the PassKit blastdoor escape details
- ASLR and PAC details
- Keep up w/ details on the related CVE:
- Google WebP related, probably other image
compression math logic
- Hardening of blastdoor further
--- Page 48 ---
Lockdown Mode
SOC-Class.com || Montance® LLC 48Copyright 2023 Montance® LLC - All Rights Reserved. All Wrongs Reversed?
--- Page 49 ---
Mitigation Offering From Apple
- Lockdown mode would thwart attack
- This was speculated by Citizen Lab, then confirmed
by Apple for this exploit
- If you are paranoid, and/or have the sort of
exposure discussed here, use Lockdown
mode
--- Page 50 ---
Recap
SOC-Class.com || Montance® LLC 50Copyright 2023 Montance® LLC - All Rights Reserved. All Wrongs Reversed?
--- Page 51 ---
Zero-click Zero Day
- Dangerous zero-click zero-day exploit discovered
- Zero-day: exploit found without a patch available
- Zero-click: user interaction not required
- Complicated exploit chain
- Bypassing blastdoor, ASLR, and PAC
- Zero day exploit code dubbed BLASTPASS by
Citizen Labs (who disclosed it to Apple)
- Lockdown mode would have prevented it
--- Page 52 ---
Your Action Items
- It’s complicated, more details in the various blog posts
- Think about the relationship between initial entry (PassKit in
this case) and ultimate exploitation (WebP in this case)
- See similar post from a couple months before BLASTPASS:
https://mgt517.com/web-gpu Safari -> GPU exploitation
by Ian Beer (iOS 16.4.1 and 16.5 fixed these)
- Chain: CVE-2023-28205, CVE-2023-28206, CVE-2023-32409
--- Page 53 ---
Christopher Crowley – Montance® LLC
- @CCrowMontance
- Slide deck (this and others)
https://montance.com/pres
(as in presentations) at my
website
- Questions?
Montance® LLC
53
Related Q&A
- Q: OverviewA: Presentation details for BLASTPASS SANS HackFest