A survey of 85 agentic-security papers finds attack work outpaces defense 3.9:1, and the action layer gets 4.7% of attention.

Shared entity: Code / Same source domain / Shared topic / Earlier coverage

Router-side injection beat every coding agent tested

Both cover Code; reported by the same outlet (arxiv.org); overlapping topics (action, agent, code, defense).

Shared entity: Code / Same source domain / Shared topic / Earlier coverage / Tension

RSIBench-Data: agents improve on their first attempt 58% of the time but finish below their own peak in 78% of runs.

Both cover Code; reported by the same outlet (arxiv.org); overlapping topics (agent, code).

Engram: a lean retrieved context beats stuffing the full history into the window.

Both cover Code; reported by the same outlet (arxiv.org); overlapping topics (agent, code).

Shared entity: Code / Same source domain / Shared topic / Earlier coverage

SimWAM scores 91.5 on NAVSIM, then discards its video model

Both cover Code; reported by the same outlet (arxiv.org); overlapping topics (action, attention, code).

Tencent: poisoned agent experience becomes a permanent skill

Both cover Code; reported by the same outlet (arxiv.org); overlapping topics (agent, attack, code).

Code-A1: Adversarial Co-Evolution Eliminates Test Suite Dependency.

Both cover Code; reported by the same outlet (arxiv.org); overlapping topics (adversarial, agent, code).

Shared entity: Action / Same source domain / Shared topic / Earlier coverage

Sleeper Cell Backdoors in Fine-Tuned Models: Poisoned LoRAs Pass All Benchmarks

Both cover Action; reported by the same outlet (arxiv.org); overlapping topics (action, agent, attack).

Shared entity: Action / Shared topic / Earlier coverage / Downstream implication

Claude Code Security: 500+ Zero-Days, Infosec Community Processing

Both cover Action; overlapping topics (action, adversarial, code); earlier Action coverage from 2026-02-24.