In the weekly report on the Linux 7.4 rc2 release candidate, Linus Torvalds stated that the wave of fixes found using AI may have reached its peak: according to eBPF maintainer Alexei Starovoitov, fixes are now entering this subsystem faster than new bug reports. This is the first measurable signal that the 'patchpocalypse' — the avalanche of machine-generated patches triggered by the Anthropic Mythos vulnerability discovery model — is finite, although the rest of the kernel has not yet reached equilibrium.

image

What happened

In the traditional weekly letter on the progress of the Linux 7.4 rc2 release candidate testing, Torvalds assessed the state of the machine-generated fix stream and suggested that their influx is close to maximum values. The basis for this was the assessment by Alexei Starovoitov, who curates the eBPF subsystem: more ready-made fixes are now arriving there than reports of new defects, meaning the AI fix pipeline has formally caught up with the flow of bug reports. In other parts of the kernel, such equilibrium has not yet been recorded. The cost of this operating mode is already visible: kernel developer Lorenzo Stoakes from Arm announced that he is reducing the volume of his reviews because, in his words, AI patch series account for 'half of what I see in review.'

Context

The press has dubbed the wave of machine-generated patches the 'patchpocalypse,' and it was triggered by the Anthropic Mythos vulnerability discovery model, presented in April 2026. Following its appearance, vendors began reporting record volumes of closed vulnerabilities: Microsoft closed 974 vulnerabilities in September 2026, compared to 622 in July and 206 in June, while Oracle reported 1,449 closed CVEs in just one month — July 2026. The Linux kernel, with over 40 million lines of code, found itself under dual pressure: models find defects faster than humans, while the throughput of human review remains the same. The bottleneck has thus shifted from the task of 'finding a bug' to the task of 'verifying and passing the flow of findings through review.'

Why this matters for the industry

For the industry, a measurable equilibrium point has emerged for the first time: eBPF became the first major kernel subsystem where the rate of incoming AI fixes exceeded the rate of new bug reports. If this mode spreads to other kernel subsystems, the backlog of machine findings will begin to clear, and the volume of mandatory patching will decrease. For now, the window between a model detecting a vulnerability and its potential exploitation by attackers is narrowing, forcing vendors to accelerate their own fix release cycles. Value in the chain is inverting: vulnerability discovery itself is being commoditized, while reviews, triage, and deduplication of machine findings become scarce resources — meaning demand is shifting from 'finding more bugs' to verification and prioritization infrastructure.

Why this matters for users

For readers, this is the first measurable signal that the 'AI patch flood' is not infinite: eBPF — the mechanism for running isolated programs inside the Linux kernel — has already reached a plateau where fixes are closed at the same rate that new bug reports are discovered. In practice, this means that Linux server and device administrators should install updates arriving with the Linux 7.4 release in a timely manner, before overloaded maintainers begin to further reduce the volume of checks. The downside is that accelerated review clearance increases the risk of missing incorrect fixes, so updating on early release candidates should still be done with caution. A simple benchmark for checking the trend: the next Linux 7.4 rc releases and Microsoft and Oracle patch Tuesdays — if the records for the number of closed CVEs stop being updated, the thesis of a passed peak will be confirmed.

What is currently unknown / limitations

The conclusion about the peak relies on the observation of a single subsystem: Starovoitov's data covers only eBPF, and there are no comparable figures for at least a second or third subsystem to interpret Torvalds' assessment. The sources lack quality metrics — the share of false positives, the share of duplicates, and the measured cost of review per patch — so the described equilibrium currently characterizes speed, not the utility of the findings. Aggregated numbers from Microsoft and Oracle mix AI findings with regular ones, making it impossible to isolate the contribution of Mythos-class models from them. Finally, there is no publication with methodology, benchmarks, and distribution of findings by criticality for the Anthropic Mythos model itself, so its effectiveness is assessed only by indirect consequences — the number of released patches.

Sources

Author

Look at AI, editorial team