CapLite derives write permissions from tool schemas and values in the user request before reading untrusted tool results. The preprint’s authors report no successful tool-goal attacks in 13,935 AgentDojo episodes across 15 open-weight models, and 29.0% utility under attack on a broader task set. Its conditional write-integrity guarantee excludes output-only goals and does not establish read confidentiality or general agent security.
AgentDojo
Dynamic environment for evaluating prompt-injection attacks and defenses on tool-using agents.
Records citing AgentDojo
Shilov and colleagues introduce Heads of Security, learned attention gates separating trusted instructions from untrusted content, tested on GPT-OSS-20B and Llama-3.3-70B-Instruct. Their preprint reports improved robustness against adaptive attacks while preserving tested utility; the mechanism depends on externally supplied trust labels and is not an end-to-end authorization system.
Amelia Kawasaki and Stefan Lee evaluate Attention Tracker and propose AttnDA, a learned alternative using attention-head features. Across five tested open models, synthetic-trained Attention Tracker averages 0.97 AUROC on deepset but 0.57 on the paper’s AgentDojo subset. Domain-matched training helps AttnDA; these are detection results rather than evidence of preventing harmful agent actions.
BRAC University researchers audit 33,000 trajectories from eleven open-weight agents (2-14B parameters) on InjecAgent and AgentDojo with a label-blind LLM judge, and report that 76.3% (AgentDojo) and 80.3% (InjecAgent) of attacks the benchmarks scored as failed were ignored by the agent, against 18.0% and 17.2% genuinely refused. Holding models and attacks fixed, they report that a system prompt permitting instructions in tool outputs raised pooled InjecAgent attack success from 8.09% at baseline to 43.05%, while a matched prohibiting prompt lowered it to 0.22%. The study uses static injections, greedy decoding and a single LLM judge, and does not test adaptive attackers or live deployments.
Researchers from Fullive-AI, Peking University, JD.com and other institutions replay AgentDojo episodes up to the step where an agent reads an injected tool output, then resample the rest many times to separate what is fixed at read time from what is decided afterwards. They report that a Qwen3-8B agent without thinking is largely decided at read time, while with thinking part of the decision forms during reasoning. Zero-shot chain-of-thought monitors on cheaper models miss much of what is already decided, while stronger monitors and a lookahead monitor that judges simulated next actions come close to the measured ceiling.
Researchers show that a hierarchical system can be compromised even when its individual agents use CaMeL, because downstream agents can reinterpret untrusted data as trusted instructions. Their multi-CaMeL protocol preserves provenance between instruction and data channels and reports no successful control-flow attacks in its benchmark tests, averaged across five model backends with no data-flow security policies configured. The result is limited to the tested attacks and hierarchical setting, with implementation limitations and reduced task success relative to an undefended system.
Huawei authors propose a plug-in layer in which fine-tuned 3B-parameter simulators, built on an uncensored model, draft a tree of the target agent's possible future actions, prune it against the actions the agent really takes, and pass the leaves to an existing detector. On AgentDojo with a Qwen3-235B agent they report attack success falling to 0% with two detectors, and on ActorAttack multi-turn jailbreaks to 0.0% at a sampling budget of 5. The attacks were benchmark attacks not designed against the defense, and no attacker aware of it was tested.
Researchers at Penn State and Purdue (arXiv v2, dated 2026-09-29; the v1 date is not in the archived text) report a way to train a reinforcement-learning prompt-injection attacker against frontier targets, where direct training finds no successful attack and so receives no reward. Their curriculum trains one attacker model against a sequence of increasingly robust targets, and they report ASR@10 of 93.8% against GPT-5.6-Luna and 45.0% against GPT-5.6-Terra on AgentDyn, where PISmith and RL-Hammer trained directly score 0%. The authors also report that the attacker transfers to six targets it was not trained on and to AgentDojo.
Li, He, Dai and Xiao (arXiv v1, 29 September 2026) propose ToolFence, an inference-time defense that compiles the authenticated user request into typed capabilities with provenance constraints on authority-sensitive arguments, enforces them with a deterministic monitor, and asks an LLM judge to grant new capability shapes rather than judge each call. On AgentDojo with Qwen3-max the authors report overall attack success of 0.20% against 21.20% undefended, with clean utility 38.90% against 42.70% undefended; a cross-session cache of approved shapes cut judge calls per task from 1.84 to 1.05 in their ablation. The attacks are six injection strategies averaged; the authors report no adaptive attacks against ToolFence and residual failures inside authorized data flows.
Nasr, Carlini, Tramèr and 11 co-authors apply gradient, reinforcement learning, search and human red-teaming attacks to 12 published defenses. Most defenses originally reported near-zero attack success, but the adaptive attacks exceed 90% success against most, and human red-teamers succeeded on every challenge in the subset of defenses they were given.
Meta open-sources LlamaFirewall, combining PromptGuard 2 (a jailbreak and injection detector), AlignmentCheck (a chain-of-thought auditor for goal hijacking) and CodeShield (static analysis of generated code). On AgentDojo, Meta reports that the combination cut attack success from 17.63% to 1.75% while utility fell from 47.73% to 42.68%.
Debenedetti and colleagues (Google, Google DeepMind, ETH Zurich) propose CaMeL, which extracts control flow from the trusted user query so untrusted data cannot change which actions run, and attaches capabilities to data to block unauthorized flows. On AgentDojo the first version reported 67% of tasks solved with provable security; the June 2025 revision, with newer models, reports 77% versus 84% for an undefended system.
NIST released the 2025 edition of its adversarial machine learning taxonomy, co-authored with the UK AI Security Institute and US AI Safety Institute staff. Unlike the 2023 edition, it includes a section on the security of agents, noting that tool-using agents are exposed to direct and indirect prompt injection and that hijacking can lead to arbitrary code execution or data exfiltration.
NIST's AI safety institute technical staff (renamed the Center for AI Standards and Innovation in June 2025) extended AgentDojo and red-teamed agents built on the upgraded Claude 3.5 Sonnet. On held-out Workspace tasks, attack success rose from 11% for the strongest baseline attack to 81% for the strongest newly developed attack, and across five injection tasks from 57% to 80% when each attack was tried 25 times. The team released an Inspect-based AgentDojo port and ran the red teaming with the UK AI Security Institute.
Debenedetti and colleagues (ETH Zurich, Invariant Labs) release AgentDojo, a dynamic environment with 97 realistic user tasks across workspace, banking, travel and Slack suites and 629 security test cases. It measures both utility and targeted attack success, and reports that existing attacks break some security properties but not all. It became the standard testbed used by CaMeL, US AISI/CAISI, LlamaFirewall and adaptive-attack studies.