AURA v0.2.0 has been released, marking a significant shift from conceptual AI safety to a hardened, deterministic security engine. Developed by kate8382, the project moves away from relying on LLMs to interpret user intent, instead implementing strict, reproducible engineering rules to detect social engineering and fraud in AI agent interactions. The release introduces 31 test suites and 393 passing tests, grounding the project in verifiable code rather than probabilistic guesswork.

From Psychological Theory to Hard Code

The core philosophy of AURA v0.2.0 is that psychological manipulation—such as fake authority claims or urgency-based bypasses—must be stopped by deterministic checks, not just prompt instructions. The project leverages the developer’s background in banking and investigative work to treat AI security like a credit committee review. Instead of trusting a prompt saying "I’m an authorized security researcher," AURA now verifies claims against local, static sources using offline cross-check adapters.

Strict Cryptographic and Policy Enforcement

Two major technical pillars in this release are the bounty-scope and pgp-verify adapters. The pgp-verify adapter enforces a strict allowlist for cryptographic keys, rejecting legacy SHA-1 and 2047-bit RSA in favor of nistP256, nistP384, nistP521, and ed25519Legacy with a minimum of 3072-bit RSA. Additionally, a new CI gate (npm run validate:keys) acts as a bouncer, failing builds if keys are expired, use unapproved curves, or exceed a maximum 2-year validity window. This ensures that trusted sources remain cryptographically sound without relying on the LLM’s ability to parse complex policy documents.

Handling Scale with Bounded Concurrency

To support real-world usage, the release introduces scripts/lib/bulk-io.ts, a worker pool for asynchronous bulk I/O operations. This module caps concurrency at 32 to prevent process crashes while preserving result ordering and enabling fail-fast cancellation on errors. This infrastructure work, led by contributor M Senanayake, addresses the practical bottlenecks of running dozens of security cases simultaneously, ensuring the engine remains stable under load.

Provenance and False-Positive Management

AURA v0.2.0 splits its test corpus into three tracks: synthetic_cases, real_cases, and benign_cases. The real_cases track mandates strict provenance metadata, including source, authors, and license, tracked in NOTICE.md. Crucially, the benign_cases track serves as a false-positive test set, ensuring the system doesn’t block legitimate educational or work-related queries. This tripartite structure allows for transparent evaluation of the engine’s precision, moving beyond simple block or allow binaries to nuanced decision-making.

Key Takeaways

  • AURA v0.2.0 replaces probabilistic LLM trust with deterministic, code-enforced security checks.
  • New CI gates enforce strict cryptographic key validity, rejecting weak algorithms and expired certificates.
  • Bounded concurrency in bulk I/O prevents crashes during high-volume security case processing.
  • The release emphasizes provenance and false-positive reduction to ensure practical usability.

The Bottom Line

This release proves that AI security isn't just about better prompts; it's about boring, reliable infrastructure. By treating AI agents like untrusted users in a banking system, AURA offers a practical, code-first approach to anti-fraud that developers can actually implement today.