CVE-2026-6420: Hardcoded attestation challenge nonce allows replay attacks

Impact

The CertificationParameters.generatechallenge() method in the push attestation protocol uses a hardcoded challenge nonce instead of generating a cryptographically random value. This removes the nonce-based replay protection from TPM quote attestation.

An attacker with root access on a monitored agent node can exploit this by stockpiling valid TPM quotes (using tpm2quote with the known nonce) before compromising the system, then replaying them to evade detection by the verifier. The push attestation timeout (~10s) constrains the generation window, but TPM throughput allows stockpiling ~50-200 quotes, enabling approximately 8-33 minutes of undetected compromise with default settings.

The attack is limited to a single agent node (AK signature binding prevents cross-agent replay). The pull-mode (legacy) attestation path is not affected.

Affected versions: >= 7.14.0, <= 7.14.1

CVSS: 6.3 Medium (CVSS:3.1/AV:L/AC:L/PR:H/UI:N/S:U/C:H/I:H/A:L)

| Metric | Value | Rationale | |---|---|---| | AV | Local | Exploitation requires local access to the agent machine (stop agent, access TPM, run replacement). The network transmission of quotes to the verifier is normal protocol operation. | | AC | Low | Deterministic attack: publicly visible nonce, standard tpm2-tools, no race conditions. | | PR | High | Root on a legitimate enrolled node is required. The vulnerability does not help gain access -- it only helps evade detection after root is obtained. No value against a machine the attacker already controls. | | UI | None | Fully automated after initial setup. | | S | Unchanged | AK signature binding confines impact to the single compromised agent. | | C | High | Compromised node continues receiving bootstrap keys, payloads, and secrets intended for trusted nodes. | | I | High | Verifier cannot distinguish a healthy system from a fully compromised one during the evasion window. | | A | Low | Only the compromised agent's revocation and incident response are suppressed; the system as a whole remains operational. |

The base score does not fully capture the operational severity: Keylime exists to detect machine compromise, so 8-33 minutes of undetected compromise is operationally critical. The fix is a one-line change and should be applied immediately regardless of the base score.

Patches

The fix restores the original random nonce generation (one-line change in keylime/models/verifier/evidence.py):

python Before (vulnerable): def generatechallenge(self, bitlength): # self.challenge = Nonce.generate(bitlength) self.challenge = bytes.fromhex("49beed365aac777dae23564f5ad0ec")

After (fixed): def generatechallenge(self, bitlength): self.challenge = Nonce.generate(bitlength)

Users should upgrade to the version containing this fix (7.14.2).

Workarounds

There is no complete workaround. The following existing mechanisms provide partial mitigation and are already active by default (no configuration needed):

1. TPM clock monotonicity check limits each distinct stockpiled quote to a single use, bounding the total evasion time. 2. Push attestation timeout (default 10s) prevents the attacker from going silent and constrains the quote generation window.

Reducing quoteinterval increases the attestation frequency but does not prevent the stockpiling attack.

References

- CWE-329: Generation of Predictable IV/Nonce (primary -- hardcoded nonce in cryptographic attestation protocol) - CWE-547: Use of Hard-Coded, Security-relevant Constants (hardcoded constant left in production code) - CWE-294: Authentication Bypass by Capture-replay (consequence -- enables replay attacks) - CWE-1241: Use of Predictable Algorithm in Random Number Generator - Introducing commit: 2bf91197 via PR #1814 - TCG TPM 2.0 Library Specification, Part 1, Section 18.4 (TPM2Quote) - IETF RATS Architecture (RFC 9334), Section 8 (Freshness)

Impact

The Keylime registrar does not enforce mutual TLS (mTLS) client certificate authentication since version 7.12.0. The registrar's TLS context is configured with ssl.CERTOPTIONAL instead of ssl.CERTREQUIRED, allowing any client to connect to protected API endpoints without presenting a valid client certificate.

Who is impacted: - All Keylime deployments running versions 7.12.0 through 7.13.0 - Environments where the registrar HTTPS port (default 8891) is network-accessible to untrusted clients

What an attacker can do: - List all registered agents (GET /v2/agents/) - enumerate the entire agent inventory - Retrieve agent details (GET /v2/agents/{uuid}) - obtain public TPM keys, certificates, and network locations (IP/port) of any agent - Delete any agent (DELETE /v2/agents/{uuid}) - remove agents from the registry, disrupting attestation services

Note: The exposed TPM data (EK, AK, certificates) consists of public keys and certificates. Private keys remain protected within TPM hardware. The HMAC secret used for challenge-response validation is stored in the database but is not exposed via the API.

Affected versions: >= 7.12.0, <= 7.13.0

Fixed versions: 7.12.2, >= 7.13.1

Patches

A patch for the affected released versions is available. It removes the line that override the configuration of ssl.verifymode, leaving the CERTREQUIRED value set by webutil.initmtls():

diff diff --git a/keylime/web/base/server.py b/keylime/web/base/server.py index 1d9a9c2..859b23a 100644 --- a/keylime/web/base/server.py +++ b/keylime/web/base/server.py @@ -2,7 +2,6 @@ import asyncio import multiprocessing from abc import ABC, abstractmethod from functools import wraps -from ssl import CERTOPTIONAL from typing import TYPECHECKING, Any, Callable, Optional

import tornado @@ -252,7 +251,6 @@ class Server(ABC): self.httpsport = config.getint(component, "tlsport", fallback=0) self.maxuploadsize = config.getint(component, "maxuploadsize", fallback=104857600) self.sslctx = webutil.initmtls(component) - self.sslctx.verifymode = CERTOPTIONAL

def get(self, pattern: str, controller: type["Controller"], action: str, allowinsecure: bool = False) -> None: """Creates a new route to handle incoming GET requests issued for paths which match the given

Users should upgrade to the patched version once it is released.

Workarounds

If upgrading is not immediately possible, apply one of the following mitigations:

1. Network isolation (Recommended)

Restrict access to the registrar HTTPS port (default 8891) using firewall rules to allow only trusted hosts (verifier, tenant):

Example using iptables iptables -A INPUT -p tcp --dport 8891 -s <verifierip> -j ACCEPT iptables -A INPUT -p tcp --dport 8891 -s <tenantip> -j ACCEPT iptables -A INPUT -p tcp --dport 8891 -j DROP

2. Reverse proxy with mTLS enforcement

Deploy a reverse proxy (nginx, HAProxy) in front of the registrar that enforces client certificate authentication:

Example nginx configuration server { listen 8891 ssl; sslcertificate /path/to/server.crt; sslcertificatekey /path/to/server.key; sslclientcertificate /path/to/ca.crt; sslverifyclient on; # Enforce client certificates

location / { proxypass https://localhost:8892; # Internal registrar port } }

A vulnerability has been identified in keylime where an attacker can exploit this flaw by registering a new agent using a different Trusted Platform Module (TPM) device but claiming an existing agent's unique identifier (UUID). This action overwrites the legitimate agent's identity, enabling the attacker to impersonate the compromised agent and potentially bypass security controls.

A flaw was found in the keylime attestation verifier, which fails to flag a device's submitted TPM quote as faulty when the quote's signature does not validate for some reason. Instead, it will only emit an error in the log without flagging the device as untrusted.

A flaw was found in keylime 5.8.1 and older. The issue in the Keylime agent and registrar code invalidates the cryptographic chain of trust from the Endorsement Key certificate to agent attestations.