Do apps launch faster in macOS Tahoe?
One seemingly widespread complaint about recent versions of macOS has been slow launch times of notarized apps even after they have cleared quarantine in their first run. Debate on the cause of this was vigorous in April and May, but was overtaken by Apple’s announcement of macOS 26 Tahoe. I therefore postponed further work on app launch and Gatekeeper until after Tahoe’s release.
This proved worthwhile, as soon after the first beta-release, it was suggested that Tahoe would accelerate the first launch of notarized apps by exempting them from a malware scan, although I have been unable to confirm that from Apple’s published information. That would make good sense: if an app’s CDHashes verify and match those of its notarization ticket, and no known malware was found when it was notarized, then it’s not plausible for it to contain malicious content since notarization was performed. This article considers what has changed in Gatekeeper in macOS 26.0.
Although Apple hasn’t yet updated it for macOS 26, its Platform Security Guide explains what Gatekeeper does in macOS Sequoia. For an app, plug-in or an installer package downloaded from outside the App Store and opened on a Mac, Gatekeeper verifies the software:
- is from an identified developer, in signature checks;
- is notarised by Apple to be free of known malicious content, in notarization ticket checks;
- hasn’t been altered, in CDHash verification;
- doesn’t contain known malicious content the first time it’s opened, regardless of how it arrived on the Mac, in a first run XProtect scan.
To assess this, I have used Calibre version 8.11.0 installed in a macOS virtual machine running on a Mac mini M4 Pro. Calibre is a large app at just over 1 GB with many components to be checked. This version has 82 items in its Frameworks folder, 20 executables in its MacOS folder, and 25 dylibs in its Plugins folder, totalling 113 dylibs and others that need to be checked.
Two VMs were prepared, one with macOS 15.7 Sequoia installed, the other with macOS 26.0 Tahoe. Both were run with 5 CPU cores, 16 GB memory, 100 GB disk, and bridged networking.
Calibre was launched three times in Tahoe, first when freshly installed and still in quarantine. Following that, the VM was shut down, started up 4 hours later and Calibre was run a second time. The VM was shut down again, and started up two days later for a third run. Log extracts were collected within a minute using LogUI.
Tahoe: First run in quarantine
This proceeded similarly to first run in Sequoia, with the exception of XProtect malware scanning. Soon after com.apple.syspolicy.exec reported that a GK (Gatekeeper) process assessment was starting, it wrote the log entrySkipping up front XProtect scan on: PST: (path: f0970f7f0ab1ff9f), (team: NTY7FVCEKP), (id: (null)), (bundle_id: (null))
That was reiterated later, shortly before the outcome of the GK evaluationSkipping XProtect scan on seen software: PST: (path: f0970f7f0ab1ff9f), (team: NTY7FVCEKP), (id: (null)), (bundle_id: (null))
There were no other log entries that suggested an XProtect malware scan had been attempted.
As is normal in the first run, notarization was checked using CloudKit, there was abundant evidence of CDHash verification, the user was prompted for their approval, and once that was complete the code was allowed to proceed.
Despite the lack of an XProtect scan, launch was as long as previous first runs in recent versions of macOS, taking 10.4 seconds from the second click to launch the app, to display of the prompt for user approval. During that period, a standard progress dialog was displayed. Entries in the log consisted of about 13,000 messages of SecKeyVerifySignature
from syspolicyd
and trustd
, with repeated errors ofSecError com.apple.securityd Malformed anchor records, not an array
Tahoe: Subsequent runs
The second and third runs, out of quarantine, were almost identical in the log, and similar to those in macOS 15.7 (below). com.apple.syspolicy.exec announced early that the code had already been evaluated, and those previous results would be used. Provenance data was found, and the evaluation completed without any CloudKit check of notarization. Times from the second click to launch the app, to loading of the app’s preferences were 0.64 and 0.66 seconds.
At various times, the kernel reported that it was considering a malware scan on one of the executables being checked, for exampleASP: considering malware scan (activeRulesVersion: 11716915239283693719 lastScanVersion: 0 chgtime: 1758884893 lastFileScanTime: 1758884893 threshold: 1758884893 pid: 949 info_path: /Applications/calibre.app/Contents/MacOS/calibre proc_path: /Applications/calibre.app/Contents/MacOS/calibre isNewerThanScan: 0 needsVersionScan: 0 is_time_eval_exempt: 0 is_scan_version_exempt: 1
Those appear new to macOS 26, and haven’t been seen in any log records from macOS 15.7 or earlier. Executables cited in those messages extended beyond components of Calibre to others that happened to be running at the time, but none was followed by any report of whether a scan was performed as a result.
Although they didn’t delay app launch, at the same time amfid
entered the path of the main executable and 112 dylibs and others to check each individually. For the dylibs and others, each was accompanied by a report from the kernel ofAMFI: constraint violation [dylib path] has entitlements but is not a main binary
Those took a period of approximately 3 seconds in each of the two runs.
Sequoia: 2nd and 3rd runs
These were remarkably similar, and in many parts identical, to those in Tahoe, except that there were no messages reporting consideration of malware scans. Times from the second click to launch the app, to loading of the app’s preferences were slightly quicker than Tahoe, at 0.59 and 0.59 seconds. Concurrent time to iterate through all the executable components was also shorter at 2.6 seconds, with identical reports of constraint violations.
Conclusions
- macOS 26.0 Tahoe does now skip reported XProtect malware scans when a notarized app is run for the first time, when in quarantine.
- Despite that, first run checks undertaken in Tahoe can still take significantly longer than in subsequent runs.
- This is the result of a vast number of SecKeyVerifySignature entries written during checks on executables.
- Tahoe also considers explicitly whether to perform malware scans on other processes being run, although none were undertaken. These are new in Tahoe.
From this evidence, I’m not convinced of any significant improvement in launch times. Please let me know whether you think that Tahoe has improved the speed of your app launches.
Previous articles:
Why some apps launch very slowly
Gone to launch