FR
live

OpenSSL discloses five flaws in QUIC, AES-GCM-SIV and CMS, and pushes fixes to 1 December

On 10 November 2026, OpenSSL publishes an advisory on five flaws — three in QUIC, two in AES-GCM-SIV, one in CMS — all rated Low, with fixes scheduled for 1 December. Map your OpenSSL 3.5, 4.0 and 4.1 applications now, starting with anything that speaks QUIC.

A wax-sealed paper envelope among a row of identical sealed envelopes, its seal cracked and slid off the flap, exposing a sliver of amber.

10 November 2026. OpenSSL publishes its November security advisory: five flaws, all rated Low, already fixed in the public repository but not yet in a release. 1 December 2026. The date versions 4.1.1, 4.0.4, and 3.5.10 will ship. Three weeks. The gap between disclosure and delivered fixes — a window in which exposure must be mapped by hand. Why it matters: the “Low” label undersells this batch, which hits QUIC, the transport on the rise, and reveals an IV-commitment forgery in AES-GCM-SIV, a bug class few teams know how to detect.

Three QUIC flaws: tomorrow’s transport is the hardest hit

QUIC is the encrypted transport underpinning HTTP/3, and it is on its way to becoming the default path for web traffic. Three of the five flaws live there, two of them memory-management defects.

CVE-2026-88822 concerns automated peer-address detection. During the handshake, the QUIC stack calls getpeername() with a buffer too small for long addresses such as IPv6: the function still returns the full address length, which is accepted as the valid input size, so the buffer ends up holding uninitialized stack bytes instead of the real address. The result is a broken connection and stack data leaking to anyone observing the outbound traffic. Versions 4.1, 4.0, and 3.5 are affected.

CVE-2026-95691 is more serious in mechanism, even if the score stays low: a retransmitted CRYPTO frame can release packet storage while the TLS record layer keeps a direct pointer into it. It is a use-after-free — a malicious QUIC server can crash the client, or extract a heap-adjacent byte through the protected handshake response. Again, 4.1, 4.0, and 3.5.

The takeaway: QUIC is young, its surface is large, and its implementations are still in the phase where memory bugs are found in bursts. Any server terminating HTTP/3 through OpenSSL is exposed to both.

Two AES-GCM-SIV flaws: when authentication goes quiet

AES-GCM-SIV is the “misuse-resistant” authenticated encryption mode, designed to forgive nonce reuse. The two flaws touching it show that even a robust mode hides traps.

CVE-2026-88823 hits the AAD (the authenticated-but-unencrypted associated data). A zero-length AAD update is treated as a reset of all accumulated AAD, instead of being ignored. An application that supplies its AAD in several fragments — one of them empty — therefore produces a message whose authentication tag does not cover the AAD supplied before the empty fragment. The consequence: data that was believed authenticated is silently left unauthenticated, and an attacker can forge the AAD of such a message.

CVE-2026-95692 is subtler and more instructive. When an application passes an IV of the wrong length through EVP_CipherInit_ex(), AES-GCM-SIV copies the IV using the expected size rather than the supplied size: a too-short IV triggers an out-of-bounds read, a too-long IV is silently truncated. The truncation breaks the binding between the IV and the computed tag: an attacker who has observed a message and its tag under one IV can produce a message that verifies under a different IV whose leading bytes match. It is not a classic forgery — the message itself is unchanged — but an IV-commitment failure. The flaw only affects applications choosing an IV size other than the standard 12 bytes, which limits severity while making the fix harder to prioritise.

The CMS hang that reaches back to OpenSSL 1.0.2

CVE-2026-107496 is the only flaw in the batch touching every still-analysed branch: 4.1, 4.0, 3.5, 3.0, 1.1.1, and 1.0.2. When CMS_encrypt() encrypts to an EC certificate whose group has order 1, random scalar generation can only produce zero, and the routine retries without limit while waiting for a non-zero scalar. The result is an infinite loop consuming a CPU core until the process is interrupted.

Exposure is conditional — the attacker must supply the recipient certificate, and the application must choose to encrypt to it. But the fix is revealing: for the 3.0, 1.1.1, and 1.0.2 branches, it is only available to premium customers. Everyone else stays exposed. One more reason not to leave end-of-life OpenSSL versions at the heart of services that handle certificates.

Why the “Low” label misleads

All five flaws are rated Low, and none is currently exploited in the wild. But the rating measures immediate exploitability, not remediation cost or bug class. This batch contains two classes that deserve particular attention.

The first is the IV-commitment failure in CVE-2026-95692: exactly the kind of bug crypto audits now hunt for, because it turns an API-calling mistake into a low-noise forgery. The second is the three-week window between disclosure and release: during that period the fixes are public — so attackers can read them — while official binaries don’t exist. Teams that wait for 1 December to act let that delay work against them.

The verdict

Map your exposure now, not on 1 December: identify what runs OpenSSL 3.5, 4.0, or 4.1 — distributions, container images, statically-linked binaries — starting with services that terminate QUIC/HTTP/3, then applications calling AES-GCM-SIV with fragmented AAD or non-standard IV sizes.

For end-of-life branches (3.0, 1.1.1, 1.0.2), CVE-2026-107496 is one more argument to migrate: the fix there is reserved for premium customers, and these versions accumulate defects without end. Schedule the 4.1.1, 4.0.4, and 3.5.10 updates for 1 December, and until then treat your QUIC boundaries as the most likely surface.

References

The cyber brief, every Tuesday

The flaws that matter and the patches to apply, in a ten-minute read.

No spam. One-click unsubscribe.
read next

On the same topic

Two HPE Juniper routers were enough to ground all US flights for four hours

On Friday 13 November 2026, two HPE Juniper routers installed by contractor L3Harris began hammering the FAA’s aging copper network at a thousand connections every four to five seconds, and a nationwide ground stop froze American aviation. It was neither a cyberattack nor, strictly speaking, a hardware failure: it is a continent-scale demonstration of what a single point of failure on legacy infrastructure costs.

← Back to the feed

Type at least two characters.

↑ ↓ navigate ↵ open esc dismiss