Adversa AI's curated GenAI security resources (October 2026)
Top GenAI security resources — October 2026
Adversa AI releases its October 2026 monthly roundup of GenAI security resources, compiling 38 resources by theme. The attack techniques section includes the reasoning code model backdoor AKRASIA (with average attack success rates up to 99.34%), InceptionRAG and Micro-Collaborative Poisoning against RAG, and activation-dilution jailbreaks that split harmful requests into subproblems to dilute safety signals. The defenses section covers CounterSteer, which suppresses indirect prompt injection using activation steering, the homomorphic encryption guardrail HE-Guardrail, and locating backdoors with sparse autoencoders. The vulnerabilities section documents an access control flaw in the LiteLLM proxy (CVE-2026-35029) and a SafeUnpickler bypass in SGLang (CVE-2026-86793, CVSS 9.8).
- Akrasia: Covert Backdoor Attacks against Reasoning Code LLMs
- Study uses XAI attribution analysis on Prompt Guard 2; synonym substitution can flip its prompt injection classification.
- Microsoft team proposes capability laundering attack: weak model decomposes tasks to extract capabilities from aligned frontier models
- InceptionRAG: Covert poisoning attacks against retrieval-augmented generation
- Inference engine fingerprinting attacks are feasible: model-driven environment discovery, exploitation, and escape