CVE-2026-48797
CriticalCVSS 9.3Exploitation Probability (EPSS)
Low risk35th percentile - higher than 35% of all known CVEs
Summary
The Backpropagate library in versions 1.1.0 and 1.1.1 has a vulnerability that allows access to the user interface without authentication. This enables attackers to upload data, trigger training runs, and access models without any security.
Risk Assessment
The lack of authentication in the user interface poses a serious risk to organizations, allowing attackers to manipulate data and perform unauthorized operations on models.
Recommendation
It is recommended to immediately upgrade to version 1.2.0. If an upgrade is not possible, run the interface without flags to restrict access to localhost and use SSH for remote access.
Related vulnerabilities
- CVE-2026-105215Critical
ZITADEL before 3.4.14 and 4.x before 4.16.2 contains an authentication bypass in the hosted Login V1 UI because the 'external account not found' registration endpoint trusts client-supplied external identity fields without a completed IdP callback. Unauthenticated attackers can submit forged IDPConfigID and ExternalUserID values to pre-create an account bound to a victim's external IdP identity, which the victim's later genuine external login then signs into.
- CVE-2026-105209Critical
ZITADEL 3.x before 3.4.15 and 4.x before 4.17.1 contains an improper authorization vulnerability: when issuing passkey or passwordless enrollment codes, it checks only the organization in the x-zitadel-orgid header, not the target user's organization. Attackers with user-write permission in one organization can obtain an enrollment code for a user in another organization on the same instance and register their own authenticator to take over that account.
- CVE-2026-105207Critical
ZITADEL 3.0.0 through 3.4.15 and 4.0.0 before 4.17.3 creates links between user accounts and external identity providers without verifying a primary factor or the caller's permission, including on identify-only Login V2 sessions and via the User Service V2 AddIDPLink endpoint. An unauthenticated attacker knowing a victim's login name can bind their own external IdP identity to the victim's account and then sign in as the victim.
- CVE-2026-103355Critical
The Unlimited Elements For Elementor plugin up to version 2.0.20 contains a blind SQL injection vulnerability. An attacker can exploit this flaw to manipulate database queries without knowing the database structure.
- CVE-2026-105135Critical
A code injection vulnerability was found in InternLM MindSearch 0.1.0 in the ExecutionAction.run function in mindsearch/agent/graph.py. The attack can be launched remotely, and exploit details have been publicly disclosed.
- CVE-2026-105134Critical
An OS command injection vulnerability was found in AhsayCBS up to version 10.3.2 in the file /rps/api/json/UpdateReceivers.do. The attack can be launched remotely, and the exploit has been published. Upgrading to version 10.3.4 resolves the issue.
- CVE-2026-88779HighActively exploited
Vulnerability in NetScaler ADC and NetScaler Gateway. This issue affects ADC: before 14.1-73.41, before 13.1-64.28, before 14.1-73.41 FIPS, and before 13.1-37.282; Gateway: before 14.1-73.41 and before 13.1-64.28.
- CVE-2026-105105Critical
Missing authentication for critical function in the ait-server telemetry and command broker in NASA-AMMOS AIT-Core through 3.1.1. An unauthenticated remote attacker with network access to the ZeroMQ message bus can inject spacecraft command data, exfiltrate command and telemetry traffic, inject forged telemetry, or disrupt the bus. The ZeroMQ broker binds XSUB and XPUB sockets to all interfaces by default without authentication.
- CVE-2026-71885Critical
In Bouncy Castle for Java before 1.86, the Messaging Layer Security (MLS, RFC 9420) implementation did not bind an X.509 credential to a LeafNode's signature_key. LeafNode.verify() checked a leaf's signature against the signature_key carried in the leaf itself, while the credential's X.509 certificate chain was stored but never parsed or validated, so the end-entity certificate's public key was never required to match signature_key as RFC 9420 sec. 5.3 requires. An attacker could present another party's certificate as its credential while signing the leaf, and the enclosing KeyPackage, with an unrelated key, and be accepted under that other party's identity through KeyPackage.verify() and the Group leaf-validation path.
- CVE-2026-92084Critical
The Beaver Builder Page Builder plugin for WordPress up to version 2.11.0.5 is vulnerable to arbitrary shortcode execution. This is due to insufficient validation before running do_shortcode, allowing unauthenticated attackers to execute arbitrary shortcodes. Exploitation requires a page with the Sidebar module containing a widget that displays attacker-controlled text, such as the Recent Comments widget with approved comments.
Original NVD description (English source)
Backpropagate is a Python library for fine-tuning large language models on a single GPU. In versions 1.1.0 and 1.1.1, the optional Reflex web UI exposes a training control plane without authentication: dataset upload, model load, training start/stop, multi-run orchestration, GGUF export, and HuggingFace Hub push. The CLI accepts two operator-facing flags intended as security controls: --auth user:pass — documented as "require HTTP Basic authentication on every request to the UI." and--share — documented as "expose the UI on a public address; requires --auth." When --auth user:pass is passed, the CLI prints Auth: enabled (user: <username>) to confirm to the operator that authentication is active, then exports BACKPROPAGATE_UI_AUTH=user:pass to the subprocess that launches the Reflex backend. The Reflex backend (backpropagate/ui_app/**) never reads BACKPROPAGATE_UI_AUTH. No authentication middleware is registered. No request-level guard runs. No WebSocket upgrade guard runs. Any client that reaches the bound port — local or remote, depending on whether --share is used — has full UI access. An inline comment at backpropagate/cli.py:1217-1218 in the v1.1.0 source documents the gap: "For Phase 1 the variable is exported but Reflex doesn't read it yet." This comment was internal-facing; the user-facing documentation (README, CHANGELOG, SHIP_GATE) advertised the contract as enforced. An attacker who reaches the bound port can read uploaded datasets, trigger arbitrary training runs against any local base models as well as read their paths, trigger HuggingFace Hub pushes and cause disk-fill DoS. This issue has been fixed in version 1.2.0. If developers cannot immediately upgrade to 1.2.0 run backprop ui with no flags so it binds to localhost, use SSH port-forwarding (ssh -L 7860:localhost:7860 <training-host>) instead of --share for remote access, and audit any host previously launched with --share, re-issuing any HF tokens used during those sessions.
Vulnerability data from NVD (NIST) · CISA KEV · EPSS

