On Thu, May 21, 2026 at 08:06:41PM +0200, Solar Designer wrote:
> Hi,
> 
> On Fri, May 22, 2026 at 02:19:42AM +0900, Hyunwoo Kim wrote:
> > With the help of several maintainers and developers, a v5 patch
> > resolving the "publicly disclosed" Dirty Frag variants other than the
> > CVE-2026-46300 (fragnesia) variant has been merged into netdev:
> > https://git.kernel.org/pub/scm/linux/kernel/git/netdev/net.git/commit/?id=48f6a5356a33dd78e7144ae1faef95ffc990aae0
> > 
> > Separately, the patch resolving CVE-2026-46300 alone has been split
> > into its own patch:
> > https://git.kernel.org/pub/scm/linux/kernel/git/netdev/net.git/commit/?id=f84eca5817390257cef78013d0112481c503b4a3
> 
> Thank you very much Hyunwoo Kim for staying on top of this and focusing
> on fixing the issues.  Such a contrast from what some others are doing.

This is a bug class I understand well, so I intend to keep an eye on it
going forward.

> 
> > This 48f6a5356a33 patch addresses four "publicly disclosed" variants:
> > 
> > 1. https://lore.kernel.org/all/agRhFtawP06hWyRa@v4bel/  (2026-05-13)
> > 2. https://lore.kernel.org/all/agSx78pXBFCdn08p@v4bel/  (2026-05-13)
> > 3. https://lore.kernel.org/all/agVpIsaSherjHTYg@sultan-box/  (2026-05-14)
> 
> Variant 3 above was found (and exploit generated) by Sultan Alsawaf, my
> colleague at CIQ, with use of "Claude Opus 4.6 (1M context)" with "hands
> tied since we didn't yet have the cybersecurity bypass." (the quotes are
> in Sultan's words)

His work was genuinely important. He precisely caught something that 
could easily have been missed.

> 
> > 4. https://github.com/v12-security/pocs/tree/main/fragnesia-5db89c99566fc  
> > (2026-05-15)
> > 
> > Note that the fourth PoC was confirmed to be blocked as well by the v3
> > fix (skb_gro_receive) [1] that resolves the third PoC,
> 
> This matches Sultan's analysis.  It may be that the rediscovery by V12
> was based on Sultan's public posting on the issue (including exploit).
> I called them out on this in their Twitter thread and got no reply.
> 
> > and the v4 [2] and v5 [3] changes address potential issues.
> > 
> > As long as the in-place path in esp remains, further variants of this
> > kind are expected to be found in the esp module. As mentioned
> > previously, I recommend keeping the mitigation in place for the time
> > being.
> 
> As a maybe better mitigation, can we somehow make in-place / zero-copy
> runtime configurable, and not only for esp?

A generic mechanism would require careful trade-off analysis, and I 
don't yet have a good idea for it.

That said, at least for the networking stack, it looks likely that 
the root cause will be addressed going forward:
https://lore.kernel.org/all/[email protected]/


Best regards,
Hyunwoo Kim

Reply via email to