Web-agent harnesses stay vulnerable across prompt and page attacks
WebTrap Farm separates defensive failures from task limits and finds harness design, not only the model, shapes outcomes.
Stories per day · 30 days
4 RSS
Relatedprompt-injection3benchmark2benchmarks2llm-agents2access-barriers1agentic-browsing1agentproofbench1agent-reliability1agents1agent-security1benchmarking1browser-agents1compile-then-execute1continual-learning1deep-search1deepseek1environment-synthesis1evaluation1
16 stories, newest first · one card per story
Stories tagged #web-agents
Baya web agent prevents prompt injections using compile-then-execute
Baya achieves 0% attack success on WASP benchmark compared to 25.5% for browser-only ReAct.
ANGLER adaptive attacker raises web agent success rate to 20.36%
ANGLER near-miss-guided framework boosts end-to-end attack success on web agents to 20.36% while lowering benign task success from 61.9% to 44.4%.
AgentHeal finds web-agent success falls sharply under access barriers
On leading LLM backbones, average task success falls from 89.4% to 43.7% when non-transient barriers block the primary path.
WebAnything rebuilds real websites into reusable agent environments
Fine-tuning Qwen3.5-9B on 5,856 trajectories lifts WebArena-Verified overall success by 13.33 points.
HaMo combines experience cards with continual learning for web agents
The framework beats harness-only and model-only adaptation on within-domain and cross-domain workflow shifts.
AgentProofBench tests reusable web-agent verifiers across 97 tasks
DeepSeek Flash and Pro keep most source programs under replay tests, while panel errors track failed deliveries.
OpenDS-World tests deep search agents on the live Web
The best of 12 frontier models scores 39.0%; humans score 21.1% across 1,000 attempts.
Intact isolates agent browsing from injected third-party page text
On 1,091 attacked pages, 12.8% of injections reached the planner versus 73.1% for a conventional planner.
Web$hell: A Unified Interface for Agentic Web Access
WebShell unifies search, browsing, and evidence reuse so LLM agents can complete long web tasks more cheaply.
WebFovea: When the Model Is Right but the Click Is Wrong -- Reliable Round Trips for Vision-Based Web Agents on Live Websites
WebFovea placed second in WebRetriever 2026 after harness fixes lifted its score from 31.0 to 57.0.
AdvSim2Real : Training Web Agents Against Adaptive Prompt Injection in a Web World Model
AdvSim2Real co-trains a 4B web agent against adaptive prompt injection, lifting completion 33.6% under an unseen adversary.
CLIFT: Conformal Self-Verification for Web Agent Training and Test-Time Scaling
CLIFT turns judge feedback into conformal self-verification for web-agent training and judge-free test-time trajectory selection.
CLIFT: Conformal Self-Verification for Web Agent Training and Test-Time Scaling
CLIFT turns judge feedback into reusable rewards and judge-free test-time selection for web agents.
HyperBrowseComp: A Multilingual and Multimodal Stress Test for Web-Browsing Agents
HyperBrowseComp stress-tests web agents with 423 hard multilingual, multimodal questions needing obscure public evidence.
RiskChainBench: A Benchmark for Obfuscated Platform Message Restoration and Evidence-Grounded Web Investigation
RiskChainBench benchmark evaluates models restoring obfuscated abuse messages and investigating risky websites, revealing execution failures as the main bottleneck.