SPLHunting query: tune before alertingPrompt injection and exfiltration in LLM app logs
Model output contains markdown images or reference links to external hosts
Indirect prompt injection makes a model emit a markdown image whose URL carries private data; when the chat UI renders it, the browser sends the data to the attacker (GitLab Duo, Mistral Le Chat, Superhuman, Google Antigravity). EchoLeak (CVE-2025-32711) used reference-style markdown to slip past link filtering in Microsoft 365 Copilot. This hunt extracts image and reference-link hosts from Bedrock model output and flags query strings carrying long encoded values. Replace the example allowlist with the hosts your app is allowed to render.
- Log source
- Amazon Bedrock model invocation logs (CloudWatch Logs or S3) in Splunk; text output logging enabled
- Techniques
- AML.T0077 LLM Response Rendering · LLM05 Improper Output Handling · LLM02 Sensitive Information Disclosure · AML.T0051 LLM Prompt Injection
- MITRE ATT&CK
- T1700AI-Enabled ExfiltrationT1698Indirect Prompt Injection
- Reviewed
- 2026-11-10
Query
index=bedrock schemaType="ModelInvocationLog"
| spath path=output.outputBodyJson output=model_output
| rex field=model_output max_match=20 "!\[[^\]]{0,200}\]\((?:https?:)?//(?<image_host>[^/)\s\"]+)(?<image_path>[^)\s\"]*)\)"
| rex field=model_output max_match=20 "\[[^\]]{1,100}\]:\s*(?:https?:)?//(?<ref_host>[^/\s\"]+)"
| eval dest_host=mvappend(image_host, ref_host)
| where isnotnull(dest_host)
| eval encoded_param=if(match(mvjoin(image_path, " "), "[?&][^=&\s]{1,40}=[A-Za-z0-9+/=_%-]{32,}"), 1, 0)
| mvexpand dest_host
| search NOT dest_host IN ("approved-cdn.example.com")
| stats count AS outputs dc(requestId) AS requests max(encoded_param) AS encoded_param values(modelId) AS models
BY dest_host identity.arn
| sort - encoded_param - outputs Expected false positives
- Applications that intentionally render images from documentation or CDN hosts (add them to the allowlist)
- Models summarising markdown documents that already contain images
Sources
- Simon Willison: markdown exfiltration incidents
- Simon Willison: EchoLeak (CVE-2025-32711) in Microsoft 365 Copilot
- AWS docs: Bedrock model invocation logging (log entry format)
Reviewed against vendor documentation, not tested in your environment. Defensive use only.