Securing autonomous AI agents: agent identity and permissions, guardrails on tool use, runtime protection, and discovering which agents an organization runs.
MarketingProductR&D
AI security posture management
Inventory of AI models, apps and data flows; misconfiguration and risk scoring; AI governance and compliance (EU AI Act, ISO 42001, NIST AI RMF).
ProductMarketing
Funding, acquisitions and partnerships
Funding rounds, acquisitions, strategic partnerships, cloud-marketplace listings, and major customer announcements.
MarketingProduct
MCP and tool-integration security
Model Context Protocol (MCP) servers, agent tool and plugin integrations, and how vendors scan, gateway, or secure them.
ProductR&D
Positioning and category claims
New taglines, category names a vendor is trying to own, comparative claims against competitors, analyst recognition, and pricing or packaging changes.
Marketing
Prompt injection and jailbreak research
New attack techniques, red-teaming results, vulnerability disclosures, and detection approaches for prompt injection and jailbreaks.
Zenity published a blog post describing a five-level maturity model (HAMMER) for securing homegrown AI agent development, emphasizing that security must be embedded in design, testing, and release pipelines rather than added late.
“Secure AI agent development is the area where the homegrown, code-first distinction matters most. When an organization buys a SaaS agent, the vendor owns the build, the guardrails, and the release pipeline. When an organization builds its own, every one of those belongs to the organization”Quoted from the original
R&D
Zenity highlights threat-modeling for agents, MCP vetting, guardrail standardization, AI-SBOM inventory, and runtime red-teaming as technical security practices, plus protecting configuration from agent self-modification.
Straiker's Head of AI announced benchmark results showing their runtime detection engine catching 98.9% of attacks with a 1.0% false alarm rate against Claude Opus 5.5, GPT-6.1 Sol, and Gemini 3.1 Pro.
“Straiker's runtime detection engine vs. Claude Opus 5.5, GPT-6.1 Sol, and Gemini 3.1 Pro as inline security judges”Quoted from the original
R&D
Their runtime detection engine uses inline security judges with LLMs; the specific integration approach and attack detection method warrants technical investigation.
HiddenLayer CEO Chris Sestito published commentary on AI security threats, focusing on how bad actors can exploit AI systems and the importance of controls in production AI deployment.
“Every single person with a bad motivation is now a capable hacker”Quoted from the original
R&D
HiddenLayer asserts that guardrails should not originate from model builders, implying architectural separation between model providers and security/control infrastructure is a core design principle.
Lasso Security published a blog post defining rogue AI agents, documenting recent incidents from Anthropic and OpenAI, and positioning their Intent Security Engine as a runtime detection and prevention platform.
“A rogue AI agent is an agent whose actions have moved outside the purpose it was built and authorized for”Quoted from the original
R&D
Lasso's Intent Security Engine uses behavioral baselines from agent configuration and interaction history, with sub-50ms inline policy enforcement at proxy/gateway layers and 300,000+ OWASP-mapped red-team payloads.
Zenity Labs disclosed AgentCorruption, a chain of security flaws in Amazon Bedrock AgentCore that allowed attackers to compromise all agents in an AWS account and region through a single prompt to a public-facing agent.
“Using a single prompt to one public-facing agent, researchers took over all AgentCore agents within the same AWS account and region, accessed private conversations and cloud credentials and persistently manipulated agent behavior”Quoted from the original
R&D
The attack exploited IMDS credential retrieval, overprivileged IAM roles shared across agents, internal agent enumeration APIs, and memory manipulation for persistence; IMDSv2 and reduced role permissions are necessary mitigations.
Straiker published a technical security analysis of a vulnerability chain in Cursor AI affecting indirect prompt injection, sandbox escape via shell builtins, and remote tunnel hijacking that enables persistent remote code execution.
“Cursor's command sandbox is bypassable with a single line. The IDE-level command parser (shouldBlockShellCommand) is blind to shell builtins like export and cd, allowing an attacker to escape the workspace scope and write to arbitrary locations in the user's home directory even with all protections enabled”Quoted from the original
R&D
The vulnerability combines three technical failures: command parser blind spots on shell builtins (export, cd, source), macOS seatbelt configured to permit home-directory writes, and an unauthenticated tunnel binary with MsgPack RPC spawn method enabling full shell access via hijacked GitHub device-code OAuth flow.