Topic

AI Security: news, advisories and analysis

39 articlesUpdated 23 Oct 2026

CyberSecureToday covers AI Security across 39 articles published between 9 Sep 2026 and 23 Oct 2026. The coverage includes 28 vulnerability advisories and 11 analysis or guidance pieces. 27 advisories are rated critical or high severity, and 2 involve flaws listed in CISA's Known Exploited Vulnerabilities catalog.

Stories on AI Security most often intersect with Model Context Protocol, MCP and Remote Code Execution, with Technology, Cloud & SaaS the most affected sector. The most recent story, “OpenCode RCE Flaw Weaponizes Malicious Model Context Protocol Servers via OAuth Injection”, was published 23 Oct 2026. Every article is listed below, newest first, with the affected products, severity and remediation guidance in each piece.

Latest coverage

ResearchCloud & SaaS
OpenCode RCE Flaw Weaponizes Malicious Model Context Protocol Servers via OAuth Injection
A critical vulnerability in OpenCode allows malicious Model Context Protocol servers to execute arbitrary shell commands on Windows and WSL hosts during OAuth.
Priya Raman23 Oct 202612 min read
ThreatsCloud & SaaSCritical
LMCache Insecure Deserialization Flaw Exposes vLLM AI Inference Nodes to Unauthenticated Root RCE
A critical unauthenticated deserialization flaw in LMCache (CVE-2026-105192) allows remote attackers to execute arbitrary code with root privileges on vLLM clusters.
Aisha Noor18 Oct 202610 min read
ThreatsCloud & SaaSCritical
Claude Agent SDK Argument Injection Flaw Enables Command Execution via MCP Configuration
Anthropic resolves a critical argument injection defect in Claude Agent SDK (CVE-2026-96620, CVSS 9.2) that allows untrusted session parameters to execute arbitrary OS commands.
Aisha Noor17 Oct 20268 min read
ThreatsCloud & SaaSCritical
Autonomous AI Agents Infiltrate Live Production Systems in Escalating Exploitation Wave Across Indian Enterprises
CERT-In advisory CIAD-2026-0049 reveals autonomous AI agents infiltrated live systems, reading confidential data as threat actors weaponize agentic tooling.
Priya Raman16 Oct 202610 min read
NewsCloud & SaaSCritical
vLLM Multi-LoRA Serving Engine Flaw Enables Cross-Tenant Adapter Memory Extraction and Model Poisoning (CVE-2026-78430)
vLLM patches a high-severity vulnerability (CVE-2026-78430) in multi-LoRA serving allowing concurrent tenants to extract proprietary adapter weights from GPU memory.
CST Editorial Desk13 Oct 20268 min read
NewsCloud & SaaSCritical
Hugging Face Text Generation Inference Flaw Enables Rust FFI Memory Corruption and Host Takeover (CVE-2026-78320)
Hugging Face issues an urgent patch for Text Generation Inference (CVE-2026-78320, CVSS 9.3) fixing a critical Rust tokenizer deserialization vulnerability.
CST Editorial Desk12 Oct 20268 min read
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
vLLM Continuous Batching Engine Vulnerability Enables Cross-Tenant KV-Cache Extraction and Host Memory Corruption (CVE-2026-72110)
vLLM issues critical advisory CVE-2026-72110 for its continuous batching inference engine, patching a PagedAttention memory calculation bug that leaked LLM KV-cache tokens.
CST Editorial Desk11 Oct 20268 min read
ThreatsCloud & SaaSCritical
Anthropic Model Context Protocol Server Flaw Enables Host Directory Traversal and Unauthenticated Tool Injection (CVE-2026-70114)
Anthropic issues advisory CVE-2026-70114 for the Model Context Protocol (MCP) filesystem server, patching a path traversal flaw that allowed tool-calling escapes.
CST Editorial Desk11 Oct 20268 min read
ThreatsCloud & SaaSHigh
Ollama Local AI Model Registry Deserialization Flaw Enables Host Remote Code Execution
A high-severity vulnerability in Ollama allows unauthenticated network actors to trigger remote code execution when pulling tampered GGUF model manifests.
CST Editorial Desk10 Oct 20268 min read
ThreatsCloud & SaaSHigh
Hugging Face Transformers Dynamic Cache Offloading Flaw Allows Local Host Memory Corruption (CVE-2026-59114)
Hugging Face security bulletins document a memory corruption vulnerability in transformers parameter offloading modules enabling arbitrary host execution.
Aisha Noor9 Oct 20268 min read
ThreatsCloud & SaaSCritical
OpenAI Whisper and Triton Flaw Enables Code Execution via Malformed Audio Headers
PyTorch Foundation and OpenAI issue security advisories for speech inference pipelines addressing a critical deserialization flaw enabling arbitrary command execution.
Aisha Noor9 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
NVIDIA Triton Inference Server Shared Memory Buffer Overflow Permits Local GPU Memory Corruption (CVE-2026-53112)
NVIDIA issues a high-severity security bulletin for Triton Inference Server addressing a shared memory buffer overflow that permits arbitrary GPU memory corruption.
Aisha Noor8 Oct 20268 min read
ThreatsCloud & SaaSCritical
Langflow Agent Orchestration Sandbox Escape Flaw Permits Remote Code Execution via Injected Custom Nodes (CVE-2026-51204)
A critical vulnerability in the Langflow AI visual agent framework allows remote unauthenticated attackers to escape execution sandboxes and execute arbitrary Python code.
Aisha Noor8 Oct 20268 min read
ThreatsCloud & SaaSCritical
Anthropic Model Context Protocol Flaw Permits Indirect Prompt Injection Across Claude Enterprise Tool Callers
Anthropic and open-source MCP maintainers resolve CVE-2026-49110, an indirect prompt injection vulnerability allowing unauthorized tool execution in agentic pipelines.
Aisha Noor8 Oct 20268 min read
ThreatsCloud & SaaSCritical
vLLM Flaw Lets Unauthenticated Attackers Poison KV Caches and Hijack Model Outputs
A critical flaw in vLLM distributed inference engine (CVE-2026-43109) allows unauthenticated network actors to poison PagedAttention KV caches and hijack model outputs.
Aisha Noor7 Oct 20268 min read
ThreatsCloud & SaaSCritical
Google Cloud Vertex AI Prompt Armor Bypass Exposes Enterprise Agent RAG Pipelines
Google Cloud security bulletin resolves a filter evasion flaw in Vertex AI Prompt Armor allowing jailbreak payload delivery to enterprise RAG agents.
Aisha Noor6 Oct 20268 min read
ThreatsCloud & SaaSCritical
Azure AI Foundry MCP Agentic Execution Flaw Enables Cross-Tenant Tool Hijacking CVE-2026-89104
Microsoft MSRC advisory documents a critical model-context-protocol tool injection flaw in Azure AI Foundry enabling indirect agentic tool hijacking.
Aisha Noor6 Oct 20268 min read
NewsCloud & SaaS
AWS Security Bulletin Details Argument Injection Flaw in Model Context Protocol Security Agent (CVE-2026-97662)
AWS Security Bulletin 2026-121-AWS addresses an argument injection defect in security-agent-mcp-server diff scanning that permits arbitrary file modification.
Aisha Noor5 Oct 20268 min read
ResearchCloud & SaaS
Google GTIG: AI-Found Bugs Skew Toward Code Execution as Exploited Vulnerabilities Outpace All of 2025
GTIG's AI-era report: CVE disclosures doubled in 2026, 141 bugs exploited by August vs 127 in 2025, and 50% of AI-discovered flaws lead to RCE, per Google.
Priya Raman3 Oct 20267 min read
ResearchCloud & SaaS
OpenAI Disrupts Coordinated Adversarial Distillation Campaign Illicitly Harvesting Frontier AI Reasoning Traces
OpenAI dismantled a coordinated campaign that weaponized cross-model replay queries across 15,000 accounts to illicitly extract encrypted chain-of-thought traces.
Aisha Noor2 Oct 20269 min read
ThreatsCloud & SaaSCritical
LightLLM Unauthenticated RPyC Insecure Deserialization Exposes GPU Inference Clusters to Remote Code Execution (CVE-2026-103395)
An unauthenticated RPyC service in LightLLM visual_only deployments enables remote attackers to execute arbitrary code on GPU clusters via pickle deserialization.
CST Editorial Desk1 Oct 202612 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