Prompt-Injection Hardening for Your AI AgentLogan Bishop
I test one existing AI agent against common prompt-injection and jailbreak techniques, then strengthen its prompt and guardrails to reduce the risk.
Scope & deliverables:
• Review of 1 AI agent's system prompt and tool/permission setup
• Testing with a set of up to 25 injection and jailbreak attempts (direct, indirect via documents/web content, data-exfiltration, role override)
• Revised system prompt with hardened instructions and boundaries
• Recommendations for guardrails outside the prompt (tool permissions, input/output filtering, human approval steps)
• Written report: what was tested, what got through, what changed, and remaining risks
• Retest of the revised prompt against the same attempts
Out of scope: Guarantees that the agent can't be bypassed, full security audit or penetration test of your app or infrastructure, writing application code for filters, multiple agents, ongoing monitoring.
What you provide: Current system prompt, description of the agent's tools and data access, a test environment or sandbox access, and examples of sensitive data or actions to protect.
Timeline: 5 days
Revisions: 1 round
Starting at$300
Duration5 days
Tags
AI agent
AI security
guardrails
jailbreak testing
LLM security
prompt injection
red teaming
system prompt
Service provided by
Logan Bishop proAnna, USA
Prompt-Injection Hardening for Your AI AgentLogan Bishop
Starting at$300
Duration5 days
Tags
AI agent
AI security
guardrails
jailbreak testing
LLM security
prompt injection
red teaming
system prompt
I test one existing AI agent against common prompt-injection and jailbreak techniques, then strengthen its prompt and guardrails to reduce the risk.
Scope & deliverables:
• Review of 1 AI agent's system prompt and tool/permission setup
• Testing with a set of up to 25 injection and jailbreak attempts (direct, indirect via documents/web content, data-exfiltration, role override)
• Revised system prompt with hardened instructions and boundaries
• Recommendations for guardrails outside the prompt (tool permissions, input/output filtering, human approval steps)
• Written report: what was tested, what got through, what changed, and remaining risks
• Retest of the revised prompt against the same attempts
Out of scope: Guarantees that the agent can't be bypassed, full security audit or penetration test of your app or infrastructure, writing application code for filters, multiple agents, ongoing monitoring.
What you provide: Current system prompt, description of the agent's tools and data access, a test environment or sandbox access, and examples of sensitive data or actions to protect.
Timeline: 5 days
Revisions: 1 round
$300