CaMeL
Research system that tracks data provenance within one agent run to keep untrusted content from steering privileged actions.
In our briefings
The claim each of the newest briefings on CaMeL would have you leave with.
On AgentDyn, the authors' benchmark of 60 open-ended tasks that make the agent replan as it goes, none of the nine defenses they ran on GPT-4o both kept the agent working and stopped attacks. Prompt Sandwiching, Spotlighting and the PromptGuard2 detector finished 55% to 63% of tasks with no attack, at or above the undefended agent's 53.3%, but left 27% to 31% of attacks succeeding. PromptGuard2 kept that work only when no attack came: under attack it finished 20.80% of tasks, because it discards a tool's whole output once it detects an injection. The ProtectAI and PIGuard detectors and the Tool Filter, CaMeL and Progent system defenses held attack success to 4.2% or less and finished 10% of tasks or fewer. DRIFT, built by three of the authors, finished the most tasks of that group with no attack, 30.0%, with 0.8% attack success, and CaMeL finished no task on any of the 12 models tested. The authors' own verdict is that none of them attain "acceptable performance for real-world deployment on AgentDyn".
None of nine prompt-injection defenses both stopped attacks and kept a GPT-4o agent working on tasks that change mid-runBI-FIND-2026-1854
Schroeder de Witt and his co-authors argue that security in multi-agent systems is "non-compositional", so "individually safe agents can compose into unsafe systems", and they make it one of their grounds for a new field. The best evidence they offer for it is cited. Jones, Dragan and Steinhardt (2024) paired Claude-3 Opus, which has strict refusal policies, with a jailbroken Llama-2 70B. The frontier model got the hard coding subtasks and the weak model the evasive phrasing, and together they produced vulnerable code 43% of the time, against under 3% for either model alone.
Each agent and each tool call can pass its check while the system fails, an Oxford-led survey arguesBI-FIND-2026-1859
By the authors' own review, most of AgentArmor's failures come from one pattern: the user tells the agent to act on instructions inside content it reads. In their failure case the user asks the agent to "read the file landlord-notices.txt and follow the instructions precisely", and the file carries an injected payment instruction. Table 1 calls this transfer execution, where the user's prompt hands control to an outside source, and the authors write that "such transfer execution is hard to deal with". Their answer is a switch with two settings. Disallowing the pattern catches more attacks; allowing it gives better utility and fewer false alarms, "since many benign runtime traces also require such transfer execution pattern." The paper does not say which setting its headline results use. We read Figure 11's bars as matching the allow setting (utility about 72%, false-positive rate about 0.07); disallowing cut attack success only from about 3% to about 2%, while utility fell to about 63% and the false-positive rate rose to about 0.38. Handling the pattern is left to future work.
AgentArmor cut attack success to 3% in its authors' tests; most misses came when a user said to follow a fileBI-FIND-2025-0279
›Cited in 18 briefings
Every briefing that discusses this term, newest first.
- None of nine prompt-injection defenses both stopped attacks and kept a GPT-4o agent working on tasks that change mid-run
Paper · · arXiv · Hao Li et al. · Feb 3, 2026
- Each agent and each tool call can pass its check while the system fails, an Oxford-led survey argues
Paper · · arXiv · Christian Schroeder de Witt · Apr 29, 2026
- AgentArmor cut attack success to 3% in its authors' tests; most misses came when a user said to follow a file
Paper · · arXiv · Peiran Wang et al. · Nov 18, 2025
- A fixed tool plan blocks prompt injection but breaks tasks; in DRIFT's tests, letting a model widen it wins most back
Paper · · arXiv · Hao Li et al. · Mar 26, 2026
- Two IETF veterans want every specification to carry guidance for AI agents, with authorization kept outside the model
Draft · · IETF Datatracker · Orie Steele et al. · Sep 11, 2026
- The IBAC paper's strict mode blocked every prompt injection, and its agent finished only a third of its tasks
Article · · ibac.dev · Jordan Potti · Mar 3, 2026
- AuthGraph checks where an agent's tool inputs come from, but trusts a poisoned source the user pointed it to
Paper · · arXiv · Peiran Wang et al. · May 26, 2026
- Checking each agent call against the user's own task stopped all 634 off-task calls the authors wrote
Paper · · arXiv · Reshabh K Sharma et al. · Mar 17, 2026
- A third-party agent talked a trusted one into unlocking the door; SEAgent blocks it, given the right labels
Paper · · arXiv · Zimo Ji et al. · Jan 17, 2026
- Refuse the risky action when its plan relies on untrusted input, a Beihang survey of agent attacks proposes
Paper · · arXiv · Guanquan Shi et al. · Dec 7, 2025
- No app, no developer: least privilege for an AI agent has to come from the task, six security researchers argue
Paper · · arXiv · Mihai Christodorescu et al. · Dec 1, 2025
- Attacks tuned to each defense got past all 12 prompt-injection and jailbreak defenses tested
Paper · · arXiv · Milad Nasr et al. · Oct 10, 2025
- A2AS promises prompt-injection defense at no latency cost, and publishes no test of it
Paper · · arXiv · Eugene Neelou et al. · Oct 8, 2025
- Tool poisoning runs through calls an agent is allowed to make; MindGuard looks for it in the model's attention
Paper · · arXiv · Zhiqiang Wang et al. · Aug 28, 2025
- Most of Progent's protection from prompt injection comes from a policy written before the agent reads anything
Paper · · arXiv · Tianneng Shi et al. · Apr 16, 2025
- Google's CaMeL stops nearly all prompt injection in tests, but agents finish fewer tasks and need written policy
Paper · · arXiv · Debenedetti · Mar 24, 2025
- OpenAI's instruction hierarchy ranks who may give orders by where the text arrived, not by who sent it
Paper · · arXiv · Eric Wallace et al. · Apr 19, 2024
- The 2023 paper that named indirect prompt injection showed an attacker can use whatever access the assistant has
Paper · · arXiv (cs.CR) · Kai Greshake et al. · Feb 23, 2023
Related terms
What appears alongside CaMeL in the same briefings.
Instruction hierarchyAgentArmorPrivilege escalationGoogle DeepMindAgentDojoProgentFIDESIsolateGPT