Adopt
Update 07/2026: Prompt injection remains unsolved - but the defense thinking has matured beyond the filter-based mitigations in our previous entry. Input validation and guardrails are still important but alone are now considered insufficient for critical guards. You need architectural defense, enforced by the harness - not by the model.
The "Lethal Trifecta"
The now-standard mental model for agent exfiltration risk. An agent is exfiltration-prone by construction when it combines all three:
- Access to private data - like sensitive files, databases, internal APIs, user context
- Exposure to untrusted content - web pages, incoming mails/issues/tickets, third-party tool results
- An external communication channel - HTTP requests, mail sending, writing to public locations
Design rule: break at least one leg per agent. Examples:
- A research agent that reads the web should not hold private credentials.
- An agent with database access should not be able to call arbitrary URLs.
Architectural defense patterns
- Plan-then-execute: the agent fixes its plan before reading untrusted content - later content may influence parameters, but cannot add new actions.
- Dual-LLM / quarantined model: a model without any tool access processes untrusted content and returns only structured, constrained results to the privileged agent.
- Capability policies in the harness:
- least-privilege tool access, allow/ask/deny rules
- read-only defaults for tools that consume external content
- separate agents/contexts per trust level instead of one agent with all permissions
- Human-in-the-loop approvals for critical actions (irreversible operations, external communication, credential use) - see the approval section in Harness Engineering. Beware approval fatigue: reserve approvals for genuinely critical actions, or users will click through them.
New attack surfaces since our last entry
- MCP-based injection: tool descriptions and tool results of third-party MCP servers are untrusted input; a malicious or compromised server can steer the agent. Restrict allowed servers and review what they expose.
- Web content targeting agents: malicious injections placed in public web pages increased measurably since late 2025 - browsing agents are now an explicit target group.
- Memory poisoning: injected content that persists into agent memory silently compromises future sessions (see ASI06 in the OWASP Agentic Top 10).
Official guidance & testing
- NSA/CISA guidance on MCP security (May 2026)
- OWASP Agentic Top 10
- Continuous red-teaming with Garak (now including agent-tool attack probes)
Adopt
Prompt injection is a critical security vulnerability where attackers manipulate LLM inputs to bypass security controls, leak data, or alter model behavior. It ranks as #1 in the OWASP LLM Top 10.
Attack Types
- Direct Injection: Malicious prompts overriding system instructions
- Indirect Injection: Attacks through contaminated data sources
- Context Manipulation: Exploiting model's context handling
- Chain Attacks: Multi-step manipulation sequences
Key Risks
- System prompt exposure
- Sensitive data leakage
- Security control bypass
- Unauthorized actions
- Output manipulation
Mitigations
- Input validation and sanitization
- Guardrails
- Output filtering and monitoring
- Permission boundaries
- Regular security testing with Garak
- Comprehensive logging