Stack layer / Threat pattern
Single-turn guardrail tests overstate your agent's real robustness: a defense worth 40 points collapses by 20 under four-turn escalation
arXiv 2607.29199
Stack layer / Threat pattern
Meta's ARCTIC Reframes Code Review as Critique for AI Diffs, Reporting Zero Defects From Self-Reviewed Changes Since Launch
arXiv 2607.29516
Stack layer / Threat pattern
AgenticRepair Hits 73% on SEC-Bench by Engineering Three Context Facets Security Engineers Assemble by Hand
arXiv 2607.29422
Stack layer / Threat pattern
CWEEP Triples Correct-Warning Rate on RTL Security Bugs Without Requiring a Security Spec
arXiv 2607.29604
Policy dependency / Stack layer
An AI agent that thought it was in a CTF published real credential-stealing malware to PyPI, and 15 machines ran it
Aikido Security
Stack layer / Threat pattern
Ransomware Response Reframed as Epidemiology, Using R0 and Re as Near-Real-Time Containment Triggers
arXiv 2607.29444
Contrast
TokTier: tokenization eats up to 64% of time-to-first-token in agent workloads, and stateful re-tokenization is 437x faster
arXiv
Policy dependency / Stack layer
Cequence Ships "Agentic Zero Trust" — One Agent-Bound Identity Across MCP, LLM and API Calls
GlobeNewswire via Yahoo Finance (corroborated by Help Net Security and VMblog)