BlackTree Security · Infrastructure · Automation · AI

Researchers Read a Linux Root Hash With Spectre Defences Still On

In a lab demonstration, Branch Target Reuse recovered a Linux root hash in three to five minutes on two recent Intel processors. It began with unprivileged local code and recovered a hash loaded into an su process, not a plaintext password or root access.

The VUSec research matters now because its paper and public repository include an end-to-end Linux cBPF exploit. The sources reviewed do not report malicious exploitation.

Old predictions survived new code

Just-in-time compilers repeatedly create, free and replace executable code. VUSec found that processors can execute the new instructions correctly while retaining an old indirect-branch prediction for the previous code.

The demonstrated attack uses classic BPF filters installed through seccomp, repeated system calls and a cache side channel. It steers transient execution through the stale prediction, then measures the traces left by data that architectural execution should not expose.

The paper reports about eight bytes per second of end-to-end leakage. After the researchers ran su root, the exploit found the su process and recovered its root password hash in an average of three minutes on Raptor Cove and five minutes on Lion Cove.

The second Linux chain also bypassed cBPF constant blinding by encoding instructions in jump offsets rather than ordinary immediate values. In that experiment, leakage reached about ten bytes per second and the hash was recovered within five minutes on both Intel systems. Neither demonstration recovered the plaintext password or established root access.

The broad finding is not a broad exploit

The project’s cross-platform results show stale-prediction behaviour on every tested processor across Intel, AMD and Arm. The complete Linux exploit ran only on the two Intel systems: Raptor Cove and Lion Cove, using Ubuntu 24.04 with kernel 6.14.0-27. The wider primitive tests also covered AMD Zen 4, Cortex-A76 and Cortex-X3. Similar stale state does not establish the same working exploit on those processors.

Firefox’s SpiderMonkey engine yielded a proof of concept for hijacking speculative control flow, but not a completed browser exploit. The researchers used Firefox 143’s standalone JavaScript shell with an instrumented measurement setup, not a stock browser. Their estimated Intel leakage rates were 36 and 62 bytes per second; a real browser chain would still need a longer transient window and a timer-resistant covert channel. GraalVM exposed the reuse condition, while its normal compilation and garbage collection removed the predictor state before a full chain completed.

Oracle merged code-cache randomisation into GraalVM’s main development branch on 19 August. According to the researchers, Mozilla considered predictor-flush mitigations and is prioritising site isolation. The research does not establish a fixed Firefox build, and a development-branch merge does not establish the first shipped GraalVM release containing that change.

The fixes target reused BPF memory

Linux addressed the demonstrated path with two linked changes:

  • CVE-2026-64508 adds the BPF mechanism for flushing indirect-branch predictions before reused JIT memory executes. The Linux CNA record identifies classic BPF as the unprivileged attack surface.
  • CVE-2026-64507 enables the x86 response when the BPF JIT and Spectre v2 mitigations are active. Its Linux CNA record says the kernel issues an Indirect Branch Prediction Barrier on BPF JIT memory reuse, while skipping that flush if the BPF dispatcher already uses a retpoline sequence.

Both Linux CNA records place the affected upstream code at Linux 5.18 and identify first fixed releases in 6.1.183, 6.6.145, 6.12.97, 6.18.39, 7.1.4 and 7.2. Distribution backports have their own package versions, so these upstream numbers are not a replacement for the installed kernel’s vendor tracker.

Update, 31 October 2026: distribution status decides the patch window

Canonical rates both issues Medium. Its first Ubuntu record and companion record, updated on 30 October, list the generic linux package as fixed in 7.0.0-31.31 for Ubuntu 26.04 LTS and 6.8.0-147.147 for 24.04 LTS. Those fixes do not clear every kernel flavour. Ubuntu 24.04’s linux-hwe-6.17 remains vulnerable; its linux-hwe-7.0 is fixed in 7.0.0-31.31~24.04.1. Ubuntu 22.04’s linux-hwe-6.8 still shows work in progress in 6.8.0-148.148~22.04.1, which is not a delivered fix.

Debian’s first tracker entry and companion entry distinguish first fixed source-package versions from packages currently listed in each suite. Both name Bookworm’s first fixed linux as 6.1.187-1 and its linux-6.12 as 6.12.100-1~deb12u1; Trixie’s first fixed linux is 6.12.100-1.

The current Bookworm security suite lists linux 6.1.190-1 and linux-6.12 6.12.113-1~deb12u1 as fixed. Bookworm’s ordinary linux 6.1.176-1 remains vulnerable. Trixie lists 6.12.107-1 as fixed and its security suite lists 6.12.113-1; Forky lists 7.2.12-1 and Sid 7.2.14-1, also fixed. Check the release, repository and source-package name as well as the version. An available fixed package does not prove a host has installed or booted it.

Shared Linux hosts deserve the first look

Prioritise shared compute, developer platforms and other Linux systems where less-trusted workloads can run locally. Record the running kernel package, then check both CVE-2026-64507 and CVE-2026-64508 against its vendor tracker.

Apply the vendor update, reboot if required and verify that the fixed kernel is running. Constant blinding is not sufficient because VUSec’s second exploit bypassed it. Canonical warns against cherry-picking the individual commits. Keep existing Spectre mitigations enabled, and use the supported kernel update for the actual installed flavour. The paper finds that IBT raises the barrier but is not sufficient on its own; race-free IBT combined with constant blinding is stronger, while predictor invalidation addresses the stale-state mechanism directly.

The Linux CNA records were published in July. The late-September story is the public research and reproducible artifacts, rather than a new CVE disclosure. The researchers’ laboratory results do not establish malicious exploitation. BlackTree’s earlier article on Tontou’s Spectre v2 timing gap covers a different technique and the same operational rule: verify the mitigation protecting the path you expose.

Sources

Leave a Reply

Your email address will not be published. Required fields are marked *