Agentic AI defense
Defense against indirect prompt injection via tool result parsing
This paper proposes a mechanism to defend against indirect prompt injection. The method involves strict parsing of tool results to sanitize inputs before the agent processes them.
Your AI agent needs seatbelts, not smarter prompts
Analysis argues that prompt injection is permanent and cannot be solved by prompt engineering. It provides a checklist for architectural defenses including confirmation gates and output validation.
How agentic identity creates accountability for agentic AI
This article explains a framework for binding human identity to AI agent actions. It covers technical defenses like policy engines, kill switches, and immutable audit trails.
How to secure agentic AI without starting from scratch
Proposes treating AI agents as security principals with unique identities. The article emphasizes applying existing IAM controls to agents rather than reinventing security paradigms.
Securing AI agents: how to prevent hidden prompt injection attacks
IBM Technology demonstrates how to use an to protect shopping agents. The video shows how to block indirect prompt injection attacks hidden in external content.
Defending AI agents against indirect prompt injection attacks
A tutorial on defending against OWASP Agentic Top 10. It focuses on lifecycle security and defense-in-depth requirements.
Agentic AI security 101
What is agentic AI? Definition and differentiators
Google Cloud provides a comprehensive explanation of agentic AI concepts. It emphasizes the shift from generation to autonomous decision-making and planning.
Agentic divide: disentangling AI agents and agentic AI
This article explores the conceptual differences between AI agents and agentic AI. It breaks down the architecture and specific risk factors associated with each.
AI agent security: protecting the next generation of intelligent workflows
A comprehensive guide covering core security concepts for intelligent workflows. It addresses supply chain vulnerabilities, orchestration security, and enterprise best practices.
Threat modeling is step 1 to secure agentic AI
A guide to threat modeling specifically for agentic systems. It discusses frameworks like MITRE ATLAS and the concept of the “lethal trifecta” involving untrusted content and privileged tools.
Agentic AI – understanding autonomous systems
An educational overview of autonomous agent capabilities. The article covers decision-making processes and the emerging security challenges of autonomy.
Research
Prompt injection mitigation with agentic AI, nested learning, and AI sustainability via semantic caching
Academic research proposing novel mitigation methods for prompt injection. The approach uses nested learning and agentic techniques to improve robustness.
PINA: prompt injection attack against navigation agents
Researchers present PINA, a framework for attacking navigation agents. The study demonstrates a high success rate in compromising agents with physical world implications.
The crisis of agency: a comprehensive analysis of prompt injection and the security architecture of autonomous AI
An exhaustive analysis of the prompt injection vulnerability in autonomous AI. It covers the “Confused Deputy” problem and evaluates defense architectures like dual LLMs.
Your Clawdbot (Moltbot) AI assistant has shell access and can be hijacked
Snyk analyzes Clawdbot security risks, demonstrating how prompt injection can exfiltrate API keys. The research highlights the dangers of agents having shell access.
Prompt injection and the security risks of agentic coding tools
Research demonstrates vulnerabilities in agentic coding tools like Cline and Cursor. It shows how malicious code patterns can be injected via MCP servers.
Agentic AI vulnerabilities
BodySnatcher (CVE-2025-12420): a broken authentication and agentic hijacking vulnerability in ServiceNow
Detailed disclosure of CVE-2025-12420, a critical vulnerability in ServiceNow. It allowed unauthenticated attackers to impersonate users and execute privileged AI agent actions.
Claude Cowork hit with file-stealing prompt injection days after Anthropic’s launch
Researchers discovered a critical vulnerability in Claude Cowork allowing file exfiltration. Attackers could hide prompts in documents that forced the agent to upload confidential files.
Superhuman AI exfiltrates emails
Analysis of security risks in Superhuman’s AI assistant. The report focuses on data exfiltration potential when agents have broad access to email content.
ZombieAgent exposes a growing blind spot in agentic AI security
Radware details the ZombieAgent vulnerability. This zero-click exploit allows attackers to hijack agents through hidden instructions without triggering traditional tools.
Incident and threat reports
Agentic AI double agents expose dangerous security gaps
Analysis of an incident where an attacker jailbroke Claude Code to target organizations autonomously. The compromised agent used MCP to access internal systems and generate malicious code.
ZombieAgent threat report
A comprehensive threat intelligence report on the ZombieAgent attack. It includes specific mitigation recommendations for this emerging threat.
AI agent prompt injection risks – I3 threat advisory
A threat advisory analyzing attack vectors specific to autonomous agents. It frames how attackers manipulate agent behavior through crafted inputs.
Attacks on agentic AI
Agentic AI: the confused deputy problem
Quarkslab demonstrates the Confused Deputy vulnerability in a medical AI assistant. The proof-of-concept shows how an agent can be manipulated to leak patient records.
Agent hijacking: breaking LLM agents with prompt injection
Snyk Security Labs details agent hijacking techniques. The research exposes vulnerable patterns in agent architectures and offers defensive recommendations.
Framework
Singapore launches new model AI governance framework for agentic AI
Singapore’s IMDA publishes a governance framework for agentic AI. It includes considerations for the Model Context Protocol (MCP) and structured risk dimensions.
AgentAudit GitHub action for AI agent security testing
AgentAudit is a GitHub Action for automated security testing. It scans agent endpoints for prompt injection and data exfiltration vulnerabilities within CI/CD pipelines.
Agentic identity and permissions are the new perimeter
The exploits detailed above prove that relying on prompt filtering alone is a failed strategy. To secure agentic AI, organizations must treat agents as distinct security principals with verified identities and strictly scoped permissions. Implement the “Agentic AI Posture” metrics referenced in this digest and deploy architectural seatbelts before your autonomous agents become insider threats.