Topic

LLM Security: news, advisories and analysis

23 articlesUpdated 12 Oct 2026

CyberSecureToday covers LLM Security across 23 articles published between 10 Sep 2026 and 12 Oct 2026. The coverage includes 17 vulnerability advisories and 6 analysis or guidance pieces. 17 advisories are rated critical or high severity, and 1 involves a flaw listed in CISA's Known Exploited Vulnerabilities catalog.

Stories on LLM Security most often intersect with AI Security, Agentic AI and Prompt Injection, with Technology, Cloud & SaaS the most affected sector. The most recent story, “LangChain LCEL Vector Index Deserialization Flaw Enables Arbitrary Code Execution via Compromised Embeddings (CVE-2026-74120)”, was published 12 Oct 2026. Every article is listed below, newest first, with the affected products, severity and remediation guidance in each piece.

Latest coverage

ThreatsCloud & SaaSCritical
LangChain LCEL Vector Index Deserialization Flaw Enables Arbitrary Code Execution via Compromised Embeddings (CVE-2026-74120)
LangChain issues advisory CVE-2026-74120 for Expression Language vectorstores, patching an unsafe deserialization flaw that allowed remote code execution.
CST Editorial Desk12 Oct 20268 min read
ThreatsCloud & SaaSCritical
AWS Bedrock Agent Runtime Flaw Enabled Persistent Jailbreaks and Tool Hijacking
AWS patches CVE-2026-68115 in Amazon Bedrock Agent Runtime where unvalidated context memory serialization allowed cross-agent jailbreaks and tool invocation hijacking.
CST Editorial Desk11 Oct 20268 min read
ThreatsCloud & SaaSCritical
LiteLLM Gateway SQL Injection Vulnerability in Vector Search Metadata Filter Allows Backend Database Takeover
A critical SQL injection flaw in LiteLLM Proxy's vector metadata filtering allows unauthenticated API consumers to execute arbitrary PostgreSQL queries.
CST Editorial Desk10 Oct 20268 min read
ThreatsCloud & SaaSCritical
LiteLLM AI Model Gateway Vulnerability Allows Unauthenticated JWT Signature Bypass and Private LLM Hijacking (CVE-2026-55102)
A critical vulnerability in LiteLLM AI Proxy allows unauthenticated actors to forge JWT tokens and hijack enterprise foundational model pipelines.
Aisha Noor9 Oct 20268 min read
ThreatsCloud & SaaSHigh
Ollama Patches Experimental Agent Mode Command Injection and Approval Bypass Flaw (CVE-2026-102697)
A security defect in Ollama experimental agent mode allowed command chaining operators to bypass confirmation gates and execute arbitrary shell commands.
Aisha Noor5 Oct 20268 min read
$3D conceptual cybernetic illustration of AI LLM GPU cluster memory pool fracturing with red glowing pointer dereference exploit indicators$
ThreatsCloud & SaaSCritical
Mooncake Distributed LLM KVCache Transfer Engine Vulnerable to Process Memory Disclosure and RCE (CVE-2026-103764)
A critical flaw in Mooncake distributed KVCache transfer engine allows unauthenticated TCP attackers to read and write arbitrary process memory, exposing LLM prompt caches.
CST Editorial Desk2 Oct 20269 min read
ThreatsCloud & SaaSCritical
GitLab AI Gateway Prompt Execution Flaw Exposes Cloud Workspaces to Remote Code Execution (CVE-2026-90970)
A critical command injection flaw in GitLab AI Gateway allows remote attackers to break LLM agent sandboxes and execute code across enterprise cloud runners.
CST Editorial Desk1 Oct 202611 min read
ThreatsCloud & SaaSHigh
LangChain Core Prompt Template Ingestion Flaw Exposes Host Systems to Path Traversal CVE-2026-34070
A high-severity path traversal vulnerability in langchain-core allows attackers processing untrusted prompt configurations to read arbitrary host filesystem data.
Daniel Okafor28 Sep 20268 min read
ThreatsCloud & SaaSCritical
Langflow Shareable Playground Flaw Enables Unauthenticated Remote Code Execution
A critical vulnerability in Langflow AI orchestration frameworks allows unauthenticated attackers to execute arbitrary code via insecure component endpoints.
Mei-Lin Chen27 Sep 20269 min read
ThreatsCloud & SaaSHigh
Ollama GGUF Parser Integer Overflow Exposes AI Inference Servers to Memory Corruption
A critical integer overflow flaw CVE-2026-86289 in the Ollama GGUF model decoder allows attackers to trigger memory corruption and remote code execution on AI inference clusters.
Priya Raman25 Sep 202616 min read
InsightsCloud & SaaS
Model Context Protocol OAuth Passthrough Flaw Exposes Autonomous AI Tool Pipelines to Silent Impersonation
An architectural vulnerability in Model Context Protocol (MCP) OAuth token passthrough allows malicious tool servers to impersonate users and hijack enterprise APIs.
Priya Raman25 Sep 202618 min read
ThreatsCloud & SaaSCritical
Azure AI Language Authentication Flaw Enables Unauthenticated Model Tampering
Microsoft mitigates maximum-severity CVSS 10.0 flaw CVE-2026-70352 in Azure AI Language Authoring, preventing unauthenticated remote tampering with enterprise AI models.
CST Editorial Desk24 Sep 202617 min read
ResearchCloud & SaaS
vLLM Multimodal Special Token Injection Triggers Remote Engine Crash and Service Denial CVE-2026-44222
High-severity flaw CVE-2026-44222 in vLLM allows remote unauthenticated users to crash distributed GPU inference clusters via crafted multimodal tokens.
CST Editorial Desk23 Sep 202615 min read
ThreatsCloud & SaaSHigh
Strands Agents Tools HTTP Request Authorization Flaw Enables Indirect Prompt Injection Credential Theft CVE-2026-18394
AWS resolves high-severity vulnerability CVE-2026-18394 in Strands Agents Tools, where indirect prompt injection forces AI agents to exfiltrate bearer credentials via malicious proxy parameters.
CST Editorial Desk22 Sep 202615 min read
ThreatsCloud & SaaSHigh
CrewAI Multi-Agent Framework Flaw CVE-2026-37007 Allows Remote Code Execution via Path Traversal Tool Poisoning
Critical path traversal vulnerability CVE-2026-37007 in CrewAI FileWriterTool allows autonomous agents to be manipulated into executing arbitrary code on host systems.
Mei-Lin Chen22 Sep 202616 min read
InsightsCloud & SaaS
Malicious MCP Servers Exploit OAuth Metadata Flaw in Rust SDK to Hijack Enterprise Agent Tokens
Critical OAuth metadata flaw CVE-2026-63127 in rmcp Rust SDK allows rogue Model Context Protocol servers to capture enterprise agent tokens and impersonate identities.
Priya Raman20 Sep 202613 min read
ThreatsCloud & SaaSHigh
Flowise SSRF Vulnerability Bypasses Network Guardrails to Steal Cloud Instance Credentials
Vulnerability CVE-2026-67620 in Flowise AI workflow builder allows attackers to bypass SSRF deny-lists and extract cloud instance metadata credentials from OCI and Alibaba Cloud.
Daniel Okafor20 Sep 202611 min read
ThreatsCloud & SaaSHigh
Amazon Bedrock AgentCore Flaw Allows Attackers to Bypass LLM Guardrails via Injected Tool-Use Payloads
AWS resolves high-severity vulnerability CVE-2026-18830 in Amazon Bedrock AgentCore allowing attackers to bypass LLM safety guardrails and trigger arbitrary tool calls.
Daniel Okafor19 Sep 202611 min read
ThreatsCloud & SaaSHigh
AWS API MCP Server Flaw Let AI Agents Bypass Security Policies
AWS Labs resolves high-severity vulnerability CVE-2026-16584 in AWS API MCP Server allowing autonomous AI agent workflows to bypass security policies upon initialization failure.
Daniel Okafor19 Sep 202611 min read
ResearchCloud & SaaS
Langflow AI Pipeline Remote Code Execution Flaws Exploited in Enterprise LLM Deployments
CISA adds critical Langflow AI vulnerabilities CVE-2026-0770 and CVE-2026-55255 to KEV catalog as threat actors weaponize agentic pipeline execution flaws.
Aisha Noor18 Sep 202611 min read
ResearchCloud & SaaS
Agentic AI Security Breakdown: How Indirect Prompt Injections Hijack Autonomous Tool-Calling Agents
Analysis of emerging frontier AI threat vectors: how indirect prompt injections in untrusted documents hijack tool-calling autonomous LLM agents.
Priya Raman17 Sep 202614 min read
ThreatsCloud & SaaSCritical
IBM Langflow Visual AI Framework Remote Code Injection Vulnerability CVE-2026-9198 Exploited in Production Environments
CISA adds CVE-2026-9198 to KEV, warning of active remote code injection attacks weaponizing visual AI workflow builder IBM Langflow in enterprise environments.
Daniel Okafor15 Sep 20269 min read
InsightsCloud & SaaS
The agentic AI security gap: 83% are deploying, 29% feel ready
Prompt injection is now OWASP’s top LLM risk with reported success rates of 50–84%. Organisations are shipping agents into production well ahead of their controls.
Priya Raman10 Sep 20268 min read