Prompt Injection Defense
Three-layer defense against prompt injection and response-blind exfiltration.
warden/sanitizer.py — structural HTML detector catching payloads the YAML regex engine misses:
- HTML comment injections (
<!-- ignore previous instructions -->) - CSS-invisible elements (
display:none,visibility:hidden,font-size:0, transparent color) aria-hiddenelement injections- Zero-width character obfuscation
Taint tracking (_TaintStore) — persists per-session injection state across hook invocations. Once a prompt injection is detected in a WebFetch response, every subsequent network call in the session is escalated to CRITICAL — closing the response-blind exfiltration gap where the attacker never needs to read a response back.
_check_cloaked_secrets_in_url — checks enrolled cloaking secrets against outbound URLs regardless of key shape, filling the gap left by static key patterns.
Two new CRITICAL policy rules:
prompt-injection-hidden— injection in HTML comments, CSS-hidden elements, aria-hidden elementssecret-in-url-params— API key shapes in outbound URL parameters (Anthropic, OpenAI, GitHub, AWS, Slack, Google, Stripe)
Expanded mini-shai-hulud IOC Coverage
New compromised packages: @opensearch-project/*, @uipath/*, PyPI mistralai==2.4.6, PyPI guardrails-ai==0.10.1
New C2: git-tanstack.com (Cloudflare-flagged phishing domain)
New payload hash: tanstack_runner.js (ce7e4199...)
New script patterns: AWS IMDS probe, HashiCorp Vault probe, GitHub GraphQL worm propagation (createCommitOnBranch), token regexes (ghp_*, npm_*), new persistence paths
Attribution: TeamPCP — same actor as March 2026 Trivy supply chain compromise. Malicious commit: tanstack/router@79ac49ee.