DARWIN: Evolving Jailbreak Adversary and Guardrail for LLM Safety Evaluation and Protection

Most existing LLM safety evaluation and defense methods follow a static formulation: jailbreak vulnerabilities are evaluated with fixed attack methods, and guardrails are trained on fixed malicious prompt datasets. However, real-world adversaries continuously evolve their capabilities and expand...

Defense against LLM Backdoors Using Critical Neuron Isolation Pruning

Large language models LLMs are vulnerable to backdoor attacks, where hidden triggers induce malicious outputs. Existing defenses generally fall into inference-time detection or training-time mitigation, but face two key limitations. First, they focus on fine-tuning-based backdoors e.g., PEFT...

PT-2026-70167

Name of the Vulnerable Software and Affected Versions n8n versions prior to 2.31.5 n8n versions prior to 2.32.1 Description A sandbox escape exists during expression evaluation. An authenticated user with permissions to create or modify workflows can use arrow-function bodies to bypass the...

CVE-2026-56746

Netty (versions 4.2.0.Final–4.2.15.Final and 4.1.0.Final–4.1.135.Final) contains a security control bypass in CorsHandler’s origin evaluation due to a logical operator error. An attacker can bypass the short-circuit protection by sending Origin: null, causing unauthorized requests to reach the ba...

CVE-2026-47391

PraisonAI is a multi-agent teams system. Prior to version 4.6.40, PraisonAI's first-party A2A server example exposes an unauthenticated A2A JSON-RPC endpoint and registers a calculateexpression tool implemented with Python eval. The example also binds to 0.0.0.0. A remote unauthenticated attacker...

CVE-2026-47391 PraisonAI's unauthenticated A2A official example can reach real LLM-driven `eval()` tool execution

PraisonAI is a multi-agent teams system. Prior to version 4.6.40, PraisonAI's first-party A2A server example exposes an unauthenticated A2A JSON-RPC endpoint and registers a calculateexpression tool implemented with Python eval. The example also binds to 0.0.0.0. A remote unauthenticated attacker...

EUVD-2026-46281

PraisonAI is a multi-agent teams system. Prior to version 4.6.40, PraisonAI's first-party A2A server example exposes an unauthenticated A2A JSON-RPC endpoint and registers a calculateexpression tool implemented with Python eval. The example also binds to 0.0.0.0. A remote unauthenticated attacker...

CVE-2026-47391

PraisonAI’s first-party A2A example exposes an unauthenticated /a2a JSON-RPC surface and registers a calculate(expression) tool implemented via Python eval(), bound to 0.0.0.0. A remote attacker can send message/send to a2a, causing the LLM to invoke the unsafe tool and potentially execute arbitr...

CVE-2026-63144: Uncontrolled Recursion

Uncontrolled Recursion CWE-674 in Elasticsearch can lead to denial of service via a specially crafted search request submitted by a low-privileged authenticated user. A user with read-level index access can submit a request that triggers unbounded recursive processing within the Elasticsearch que...

promptfoo v0.121.19

Promptfoo: LLM evals & red teaming promptfoo is a CLI and library for evaluating and red-teaming LLM apps. Stop the trial-and-error approach - start shipping secure, reliable AI apps. Website · Getting Started · Red Teaming · Documentation · Discord Promptfoo is now part of OpenAI. Promptfoo...

CVE-2026-40187 Authenticated RCE via Malicious eTemplate Upload in EGroupware

In egroupware version 26.0 and earlier, an authenticated administrator can achieve OS-level Remote Code Execution RCE by uploading a malicious eTemplate XML file .xet to the VFS /etemplates mount. The Widget::expandname method passes template widget attribute values directly into a PHP eval call...

redteam-ai-benchmark — Updated!

Red Team AI Benchmark Russian version: README.ru.md Red Team AI Benchmark is a CLI model-evaluation benchmark. It measures how LLMs understand and respond to red-team questions and security scenarios; it is not a tool for carrying out those activities. Version 2 uses a rubric-based dataset instea...

EUVD-2026-45926

SurrealDB before 3.1.0 evaluates user-supplied WHERE clauses in SELECT statements and SET/MERGE/CONTENT/PATCH clauses in UPDATE, UPSERT, INSERT ON DUPLICATE KEY UPDATE, and RELATE update-variant statements against full record data before enforcing PERMISSIONS FOR SELECT WHERE restrictions. An...

CVE-2026-63755

SurrealDB before 3.1.0 evaluates user-supplied WHERE clauses in SELECT statements and SET/MERGE/CONTENT/PATCH clauses in UPDATE, UPSERT, INSERT ON DUPLICATE KEY UPDATE, and RELATE update-variant statements against full record data before enforcing PERMISSIONS FOR SELECT WHERE restrictions. An...

CVE-2026-63754 SurrealDB before 3.1.0 Denial of Service via LIVE Query

SurrealDB versions before 3.1.0 contain a denial of service vulnerability where malicious LIVE queries with WHERE clauses that evaluate to errors cause all CREATE, UPDATE, and DELETE operations on the watched table to fail. An authenticated user with only select permission can prevent write...

CVE-2026-63743 SurrealDB before 3.1.0 Port-Specific Deny Rule Bypass via HTTP Redirect

SurrealDB before 3.1.0 contains a capability bypass vulnerability in HTTP redirect handling that allows authenticated users to circumvent port-scoped --deny-net rules. Attackers can chain an HTTP redirect from an allowed hostname to a denied host:port combination, and the redirect is followed...

Adversarial Robustness of Phishing Email Detection: A Comparative Study of TF-IDF + Logistic Regression and Fine-Tuned DistilBERT

Phishing emails remain one of the most persistent cybersecurity threats, and machine-learning classifiers are widely used to detect them. Most reported detection accuracies, however, are measured on clean, in-distribution test data rather than on emails deliberately altered to evade detection. Th...

PT-2026-61559

Name of the Vulnerable Software and Affected Versions SurrealDB versions prior to 3.1.0 Description An issue in HTTP redirect handling allows authenticated users to bypass capability restrictions. Specifically, attackers can circumvent port-scoped --deny-net rules by chaining an HTTP redirect fro...

Measuring and Evaluating the Performance of Generative AI Models for Scam Detection

Online scams continue to cause substantial financial and personal harm. As a result, detection systems based on Large Language Models LLMs have been integrated into security products ranging from email gateways and browser extensions to fraud-monitoring dashboards. As this adoption accelerates, a...

How Jailbreak Attacks Inform Safety Alignment: A Defender-Centric, Shapley-Based Evaluation of Jailbreak Contributions

Jailbreak attacks on large language models are usually evaluated by attacker-centric metrics such as attack success rate ASR, yet an attack that breaks a model is not necessarily useful for improving its safety. We propose a defender-centric view of jailbreak evaluation, where attacks are evaluat...