Skip to main content
Search

# Prompt Injection

31 related articles

Subscribe to updates
  1. IBM ships LLMON markup and Granite 4.1 preview models: separating instructions from data leaves most indirect prompt injections unexecuted

    IBM released LLMON (LLM Object Notation), a markup language, together with two experimental preview models trained from Granite 4.1 3B and 8B base checkpoints, granite-4.1-3b-LLMON-preview and granite-4.1-8b-LLMON-preview, which use explicit tags to separate instructions, data and requirements. On the OPI benchmark the LLMON models avoided 47% more attacks than the best frontier model tested and 95% more than the best open-source model, while general capability stayed on par with Granite 4.1 Instruct. IBM describes the release as an experimental preview for exploration and evaluation, not for production.

  2. arXiv

    SecureGARD puts a semantic gate on tool calls: zero observed attack success across three benchmarks and two backbones

    Tool-using LLM agents turn model output into real actions, making indirect prompt injection an execution-security problem. SecureGARD inserts an executable ontology between planning and tool execution, grounding heterogeneous tool calls into shared semantic actions and judging each as ready, waiting, or forbidden against the current task state. Under benign and indirect-injection settings on AgentDojo, ASB, and AgentDyn, the observed attack success rate is 0% on both backbones, with higher task utility under attack than the compared baselines.

  3. arXiv

    Agent swarms learn adaptive prompt injection: even the strongest defender, Gemini 4 Argon, resists only 66% of attacks

    After prompt injection attacks neared saturation on public benchmarks, an optimizer that runs a swarm of collaborating agents to search for attack payloads pushed success back up. The study built 100 synthetic tool-use scenarios whose attack goal is a random tool call unrelated to the user task; against the strongest attacker, Gemini 4 Argon, the most robust defender, Gemini 4 Argon, resisted only 66% of attacks at the full query budget and 87% at 150 queries.

  4. arXiv

    Web agents stop deciding mid-page: Baya compiles the task first, cutting injection success to 0%

    Web agents usually follow ReAct, choosing each action from page observations that mix trusted and untrusted content, so injected text can rewrite their control flow. Baya compiles the user request into a Python program that calls a Trusted Website API before touching the live site; the planning model never sees live page content, and only the program reads it at runtime. On 581 WebArena Verified tasks Baya reaches 77.8% aggregate utility versus 55.1% for browser-only ReAct and 66.7% for API-enabled ReAct, and on WASP it records a 0% attack success rate against 16.2% and 25.5% for those baselines.

  5. Gemini

    Gemini 4 Argon opens to more cyber defenders: simulated attack scenarios solved rise from five to all twelve

    Gemini 4 Argon is rolling out to a broader set of cyber defenders after a limited Fairwind window, and it solved all twelve end-to-end simulated cyberattack scenarios, up from five in the previous generation.

  6. arXiv

    Coding agents with robot arms: ArmedAgent finds up to 94.29% of tasks end in an unsafe state

    Frontier models' spatial understanding now lets coding agents control a real robot arm zero-shot and end-to-end, with no task-specific demonstrations or retraining, so risk crosses from the digital world into the physical one. ArmedAgent sorts these risks by source into benign instructions that turn unsafe, malicious use, and indirect prompt injection, across 36 tasks in 5 scenarios, evaluated in simulation and validated on a physical arm. Testing GPT-6-Astra, GPT-6-Sol and GPT-6.1-Terra under Codex, overall attack success ranges from 72.22% to 94.29%, and refusal rate is 0% for every model in every category.

  7. arXiv

    A self-evolving trading agent rewrites its own prompts, but Steward locks risk rules beyond the type system and lost just 0.3% in a 24.6% sector drawdown

    Autonomous investing agents must learn from outcomes without rewriting their own risk rules. Steward separates editable prompts from an invariant core: nine role-specialised agents decide jointly, code grades every trade nightly, and a quarterly reflector proposes only typed prompt edits that a deterministic gate applies reversibly. Over 102 trading days and 1,632 model calls in a paper account, the portfolio lost 0.3% while the semiconductor sector fell 24.6%, and no injected instruction was followed.

  8. Right tool, swapped argument: BiCP cuts confirmed injected attacks from 83/682 to 0/682 on a 922-case fixed replay

    Tool-using language-model agents can pick the intended tool yet execute an injected argument, and existing counterfactual guards lose source sensitivity when they score with a reasoning prefix that already encodes the injected decision. BiCP removes the current-step reasoning prefix before scoring, probes both the tool-name target and the complete-call target, and routes proposals by operation risk. On a 922-case fixed replay, the full method reduces confirmed InjecAgent attacks from 83/682 to 0/682 and unsafe conditional executions from 33/240 to 0/240, while benign task success rises from 66/120 to 78/120.

  9. OpenRouter Blog

    OpenRouter puts the code sandbox inside the API request: any model can run shell server-side, with network off by default

    OpenRouter has released two server-side code execution tools, openrouter:shell and openrouter:bash, which let a model run commands inside a platform-hosted isolated container during the API request instead of the developer operating a sandbox. Previously OpenAI, Anthropic and Google each offered this only for their own models, while self-hosting meant choosing an image, provisioning compute and owning the security boundary. Both tools are in beta, outbound network is off by default, sandbox time is billed at $0.0001 per second, and a new or sleeping container carries a 30-second minimum.

  10. arXiv

    LeanGuard blocks forbidden tool calls before they run: direct prompt-injection success on DTap falls from 57.9% to 5.9%

    Tool-using language-model agents often execute actions their domain policies forbid, and LeanGuard places policies written in Lean 4 between the model and its tools so that only admitted calls execute. On τ 2-bench, unguarded runs of five Qwen models executed at least 33 calls the Lean policy forbids, while LeanGuard executed none; on DTap, direct prompt-injection success fell from 57.9% to 5.9% and indirect success from 22.3% to 3.8%. Denials with feedback did not significantly change task success, and the read-only deployment blocks many benign writes.