Salt Introduces Prompt Security with new AI-DR Solution
Language

Security Intelligence

Hub API and Agentic
Threat Monitoring

0
Stories tracked
1
Days covered
June 20, 2026
Oct 12, 2026
Latest brief

Threat archive

Thank you! Your submission has been received!
Oops! Something went wrong while submitting the form.
October 12, 2026

AWS Releases Strands Box, an Open-Source Policy-Enforcing Sandbox for AI Agents

AWS open-sourced Strands Box, a developer-preview sandbox that places agents in a kernel-enforced macOS Seatbelt sandbox and routes shell commands, Python code, MCP tool calls and network traffic through four policy checkpoints evaluated by the Dogwood engine. A shared event history lets rules link actions across tools — for example blocking outbound HTTP after a read from a customer-data directory, or capping a payments API at $100 a day. A network gateway swaps placeholder tokens for real API keys so agents never hold secrets.

The design treats the agent's API and MCP egress as the enforcement point — exactly where agent abuse must be governed. It also highlights that policy coverage gaps remain for directly granted resources.

Agentic/MCP
Read full story
October 12, 2026

Attackers Exploit CVSS 10.0 Bricksforge WordPress Flaw for Unauthenticated RCE via REST API Endpoint

Patchstack observed active exploitation beginning October 7, 2026 of CVE-2026-85097 in the Bricksforge WordPress plugin. Attackers fetch a nonce from the unauthenticated bricksforge_regenerate_nonce AJAX endpoint, upload a GIF/PHP polyglot, then submit a form to the /wp-json/bricksforge/v1/form_submit REST endpoint with a crafted temporaryFileUploads parameter whose url field ends in .php. Telemetry recorded 63 unique source IPs testing extension and encoding variants.

A textbook API authorization and input-validation failure: the endpoint trusts client-supplied metadata. Runtime API monitoring catches this parameter abuse that static scanning misses.

API Security
Read full story
October 12, 2026

Personal AI Agents Are Already Authenticating Into Corporate Systems via OAuth, API Keys and Saved Passwords

The Hacker News examined how employee-owned agents such as OpenClaw, Meta's Muse, OpenAI's Dots, GrokBot and Instinct reach company systems without IT approval — through user-granted OAuth consent to Workspace, Microsoft 365 and Slack, saved passwords in agent browsers, long-lived API keys and personal access tokens, and local agents on employee laptops. Vendor-side governance only applies to managed corporate deployments, leaving truly personal agents invisible in admin consoles.

Shadow agents inherit their owner's full API access while appearing as ordinary user traffic. Discovering OAuth grants, unknown API keys and unattributed sessions is the only practical control plane.

Agentic/MCP
Read full story
October 9, 2026

GhostAction Campaign Plants Credential-Stealing GitHub Actions Workflows in Tens of Thousands of Repositories

Attackers compromised maintainer accounts — including those behind pyxel and Uber's athenadriver — to push fake "Security Audit" workflows into over 340 repositories, with Socket counting more than 500 accounts committing the workflow across tens of thousands of repos since October 7, 2026. The workflows scan the working tree and full git history for 13 credential patterns and exfiltrate AWS keys, GitHub/GitLab tokens, and Anthropic, OpenAI and OpenRouter API keys over plain HTTP.

Harvested LLM and cloud API keys become direct API abuse at scale. Monitoring for anomalous use of stolen keys is the control that catches exploitation after credentials leave CI.

Breach
Read full story
October 9, 2026

Attackers Chain AhsayCBS Auth Bypass and Command Injection Zero-Days to Drop Web Shells and XMRig Miners

Huntress observed exploitation starting October 7, 2026 of CVE-2026-105133, an improper authentication flaw in the checkSysPwd() function of AhsayCBS's API component, chained with CVE-2026-105134 (CVSS 9.3), an OS command injection in the Replication Receiver. Attackers achieved RCE, dropped web shells and XMRig miners disguised as edge.exe, and deployed an apparently AI-assisted PowerShell script that kills Windows Task Manager. Version 10.3.4 remains vulnerable, making these effective zero-days.

An authentication bypass in an API component chained to command injection shows how weak API auth becomes full server compromise. Restricting and monitoring management API exposure is the immediate mitigation.

API Security
Read full story
October 9, 2026

Exposed NVIDIA DCGM Exporters Leak Telemetry From Over 12,000 GPUs via Unauthenticated Endpoints

Researchers disclosed CVE-2026-47483, a high-severity flaw in NVIDIA's DCGM Exporter GPU monitoring tool, where unauthenticated attackers can flood profiling endpoints to exhaust resources and crash the service, potentially degrading AI workloads on the same host. Scans between March and May 2026 found roughly 2,100 internet-exposed hosts running the exporter without authentication, leaking telemetry from more than 12,000 GPUs. NVIDIA fixed the issue in version 4.8.2.

Unauthenticated monitoring endpoints in AI infrastructure leak operational data and offer a denial-of-service path. API discovery across AI compute estates is needed to find these exposed services.

LLM
Read full story
October 8, 2026

ARTEX Agentic Pentesting Tool and LLMs Used to Steal Data From South Korean Banks

CrowdStrike detailed a campaign from late September to early October 2026 that used ARTEX, an open-source Chinese LLM multi-agent autonomous pentesting tool, alongside LLMs including DeepSeek, GLM and Grok to exfiltrate data from South Korean financial firms such as Shinhan Bank and Yegaram Savings Bank. Researchers found exposed Claude Code session histories and ARTEX configs on a Hong Kong IP. The developer has since taken ARTEX closed source.

Autonomous agents now probe authentication surfaces and APIs at machine speed. Behavioral detection of abnormal API access patterns matters more than static rules against AI-driven reconnaissance.

Agentic/MCP
Read full story
October 8, 2026

Tensorlake npm SDK Compromised to Deliver Shai-Hulud Credential-Stealing Worm

Version 0.5.144 of the "tensorlake" npm SDK — used for Tensorlake AI agent sandboxes and cloud services — was compromised in a ChainDrop/Shai-Hulud supply chain attack. Socket found an obfuscated preinstall worm that harvests npm, GitHub, AWS, Kubernetes, Vault and SSH credentials, plus MCP and AI coding tool config files for Claude, Cursor, Kiro, Windsurf and Zed. It establishes persistence, executes remotely supplied code, and republishes itself via the victim's publishing identity.

Stolen API tokens and MCP configuration files turn one poisoned dependency into broad API and agent-infrastructure access. Teams need visibility into which APIs and agent credentials are reachable after such compromises.

Breach
Read full story