Learning to Inject

Key findings

• RL-trained attacker models find injections that transfer across agents and defences - Automated attackers outperform handcrafted and optimisation-based baselines

Sources

• arxiv.orghttps://arxiv.org/abs/2602.05746