Part 2: Apple, the device that defends itself
Part 2 of 6 in a deep dive on machine learning that predicts attacks and stops them early, with Apple as the main case. Start with Part 1 if you are new to the series.
Random tags in silicon
On September 9, 2025, Apple announced Memory Integrity Enforcement, or MIE, alongside the iPhone 17 and iPhone Air. Apple says its chip and operating system teams spent five years on it (Apple Security Research). MIE targets memory corruption. According to Apple, the mercenary spyware chains known to target iOS, Windows and Android all exploit memory safety bugs, and those bugs are interchangeable, so attackers swap in a new one when an old one is fixed.
The mechanism is easy to describe. Every memory allocation gets a secret tag, and the hardware checks the tag on every access. Neighboring allocations get different tags, so a write that runs past the end of one buffer into the next one fails. When memory is freed and reused, it gets a new tag, so a stale pointer left over from a use-after-free bug fails as well.
Randomness does the protecting here. Apple writes that its allocators assign random tags, and that it re-seeds the pseudo-random generator behind them often, so an attacker cannot predict the tag the next allocation will get. Apple also designed its chips so that tag values cannot affect speculative execution, after researchers showed with TikTag and StickyTags that the MTE implementation on Google's Pixel phones leaked tags through timing and could be bypassed.
Apple's offensive research team attacked the design from 2020 to 2025, first on paper, then in simulation and then on prototype chips. It tried to rebuild six real-world exploit chains against MIE and could not get any of them to work, even after swapping in new bugs. MIE now covers the kernel and more than 70 userland processes on A19 and M5 chips and later, and developers can turn its tag checking on for their own apps in Xcode.
MIE uses no machine learning. I include it because it works in the opposite direction from a model. A model tries to predict the attacker, and MIE makes sure the attacker cannot predict the system. In my first deep dive I looked at keys broken by weak randomness, from Debian's OpenSSL bug in 2008 to the Coldcard seeds drained in July 2026. MIE is a case where good randomness is the defense.
Shrinking the target, paying the attackers
MIE protects everyone by default. Lockdown Mode is for the few people who face mercenary spyware. It switches off features that zero-click exploits have used, such as most message attachment types and some web technologies. It has a track record. In April 2023 Citizen Lab found that it blocked an NSO Group exploit that went through HomeKit, and the phone even warned its owner. In September 2023 Apple told Citizen Lab that the mode blocks BLASTPASS, an NSO zero-click chain delivered through images in iMessage (Citizen Lab). That December Apple said it was not aware of any successful attack on a device running Lockdown Mode (TechCrunch).
Apple also pays people to find what its own team missed. In October 2025 it doubled its top bounty to $2 million for a zero-click exploit chain comparable to mercenary spyware, with bonuses for Lockdown Mode bypasses that can take a payout above $5 million. Apple says it has paid $35 million to more than 800 researchers since 2020 (Security Affairs). For the servers that answer Apple Intelligence requests, Private Cloud Compute, Apple released a virtual research environment and source code for key parts in October 2024 and offers up to $1 million for remote code execution (Apple Security Research).
A bounty puts a price on finding the bug before a spyware vendor does. A bypass reported to Apple gets fixed before an attacker can use it, and the extra bonus for Lockdown Mode bypasses tells researchers where Apple wants the pressure.
Apple's device layers, by year
Pointer Authentication Codes
Apple ships PAC in the A12 Bionic, the first in the industry, to protect code flow when memory is already corrupted (Apple).
Threat notifications begin
Apple starts warning people it believes were individually targeted by mercenary spyware. By 2026 it has notified users in more than 150 countries (Apple Support).
Lockdown Mode
iOS 16 adds an optional mode for people at high risk of targeted attacks. It switches off parts of the system that zero-click exploits rely on.
Lockdown Mode blocks real attacks
Citizen Lab and Apple confirm that Lockdown Mode blocked an NSO Group exploit against HomeKit and the BLASTPASS zero-click chain, and in December Apple says it is not aware of anyone using the mode being hacked (TechCrunch).
Private Cloud Compute opened to researchers
Apple publishes a virtual research environment and source code for the servers that run Apple Intelligence requests, with bounties up to $1 million (Apple Security Research).
Stolen Device Protection
iOS 17.3 adds extra checks when an iPhone is away from familiar places such as home or work (Apple Support).
MIE and on-device scam filters
MIE ships with the iPhone 17. iOS 26 adds a spam filter in Messages and Call Screening for unknown numbers, both running on the device (Apple Newsroom).
A $2 million bounty
Apple doubles its top bounty to $2 million for zero-click exploit chains, with bonuses for Lockdown Mode bypasses that can push a payout above $5 million (Security Affairs).
Alerts on the iPhone itself
Apple now delivers threat notifications on the iPhone as well as by email. Waves went to users in 110 countries on August 13 (TechRepublic) and 99 countries on October 30 (CyberScoop).
Models that run on the phone
Apple's on-device features share one design rule. The model runs where the data is, and at most a result leaves the phone.
| Feature | What it checks | What leaves the device |
|---|---|---|
| Messages spam filter (iOS 26) | Sorts spam and junk into their own folder and turns off the links in them | Nothing. Apple says the sorting happens on the device |
| Call Screening (iOS 26) | Asks an unknown caller for a name and a reason before the phone rings | Nothing. It runs on the device |
| Stolen Device Protection | Whether the iPhone is at a familiar place, learned from Significant Locations | Nothing. Away from those places, key changes need Face ID, an hour's wait and Face ID again |
| Threat notifications | Apple's threat intelligence detects targeted spyware attacks | An alert on the iPhone and an email to the person targeted |
Stolen Device Protection shows how small a useful model can be. The phone knows where its owner usually uses it. A thief who has watched the owner type the passcode can still unlock the phone, but cannot change the Apple Account password, the device passcode or the Face ID setup without Face ID, a one-hour wait and Face ID again. Arriving at a familiar place can end the wait early (Apple Support).
Threat notifications sit at the other end, with the most serious attackers. Apple calls them high-confidence alerts that a person was targeted individually (Apple Support), and it recommends that people who receive one turn on Lockdown Mode (TechRepublic). Apple does not say how it detects these attacks. The notifications show the result: a defender with a view across all of Apple's devices can find and warn the few people being targeted.
What the device teaches
Remove whole bug classes
MIE makes most memory corruption unusable, so the models behind it face fewer attacks. Random tags keep attackers from learning the system.
Score where the data lives
Spam filtering, call screening and the location check run on the phone. Threat notifications reach the person directly, and private content stays on the device.
I would copy this layering anywhere. The hardware removes bug classes, small models on the device catch what reaches the person, and servers see only verdicts. Part 3 moves from one device to the whole software ecosystem, and to models that predict which bugs attackers will use next.
Continue to Part 3: Apple, the store