LoLLMs CVE-2026-1114

Remote unauthenticated exploitation against default configurations of LoLLMs 2.1.0. The only prerequisite is network reachability to the LoLLMs web interface or API endpoints-no authentication, special privileges, or user interaction required per CVSS vector AV:N/AC:L/PR:N/UI:N. The weak JWT secret appears to be a default or hardcoded value in version 2.1.0, meaning all unpatched instances share the same vulnerability without requiring any non-default configuration. Exploitation is a two-phase attack: phase one (offline brute-force) requires obtaining any valid JWT token, which can be trivially acquired by creating a free user account if registration is enabled, or intercepting network traffic if the application lacks HTTPS enforcement. Phase two (token forgery and replay) requires only the recovered secret and basic JWT manipulation tools. Limiting factors: exploitation requires technical knowledge of JWT internals and access to brute-force tooling (hashcat, John the Ripper), which slightly raises the skill floor beyond script-kiddie level. The offline attack phase leaves no logs on the target system, making detection impossible until forged tokens are used. Deployments behind VPNs, IP allowlists, or with user registration disabled reduce initial token acquisition opportunities but do not prevent exploitation by authenticated low-privilege users or network-adjacent attackers.

CVSS 9.8 (Critical) reflects network attack vector, low complexity, no privileges or user interaction required, and high impact across confidentiality, integrity, and availability-textbook authentication bypass scenario. However, real-world risk factors present a mixed picture. EPSS score of 0.04% (14th percentile) indicates very low observed exploitation probability in the wild, suggesting this vulnerability affects a niche product with limited attacker interest or that exploitation requires specialized knowledge of the LoLLMs deployment environment. SSVC assessment confirms proof-of-concept availability and classifies the attack as automatable with total technical impact, which would normally elevate priority, but the exploitation=poc status (not 'active') and low EPSS suggest targeted rather than opportunistic threats. No CISA KEV listing confirms this is not being exploited at scale. Primary risk is to organizations using LoLLMs 2.1.0 in production environments where administrative access provides lateral movement or data exfiltration opportunities. The offline nature of the attack means defenders cannot detect brute-forcing in logs, making this a silent privilege escalation vector once an attacker obtains any valid JWT (even a low-privilege one) to use as a template. Patch availability in 2.2.0 significantly reduces risk for actively maintained deployments, but unmaintained or containerized instances may remain vulnerable indefinitely.

An attacker with network access to a LoLLMs 2.1.0 instance first captures any JWT token by registering a low-privilege account or intercepting traffic. Using the captured token as a template, the attacker extracts the HMAC signature and header/payload, then launches an offline brute-force attack using hashcat with common password dictionaries and weak key patterns against the JWT signature. Within hours (depending on secret complexity), the attacker recovers the signing secret. They then modify the JWT payload to set administrative flags (e.g., 'role':'admin', 'isAdmin':true), resign the token using the cracked secret and HS256 algorithm, and replay the forged token in API requests to restricted administrative endpoints. The application validates the signature successfully and grants full administrative access, allowing the attacker to extract sensitive data, modify system configurations, create backdoor accounts, or deploy malicious payloads through LoLLMs model management features. Proof-of-concept code demonstrating JWT cracking and forgery is publicly available per SSVC exploitation status.