Google's A2A protocol provides a secure communication framework for AI agents, but it has critical limitations when handling highly sensitive information such as payment credentials and identity documents. These gaps increase the risk of unintended harms, including unauthorized disclosure, privilege escalation, and misuse of private data in generative multi-agent environments. In this paper, we identify key weaknesses in A2A: insufficient token lifetime control, lack of strong customer authentication, overbroad access scopes, and missing consent flows. We propose a modular, interceptor-based architecture to address these weaknesses. Our refinements introduce a Zero-Trust Interceptor that enforces explicit consent orchestration, ephemeral scoped tokens, and direct user-to-service data channels to minimize exposure across time, context, and topology. A component-level evaluation of the interceptor's deterministic enforcement logic, conducted across 5,000 protocol-enforcement trials, demonstrates that the enhanced protocol structurally eliminates the sensitive-data leakage paths covered by the tested checks, yielding a 0% structural success rate under the stated enforcement assumptions and keeping enforcement overhead at the sub-millisecond level. Comparative analysis highlights the advantages of our deterministic, protocol-level enforcement over both the original A2A specification and stochastic model-based safeguards. These contributions establish a practical, statistically validated path for evolving A2A into a privacy-preserving framework for multi-agent systems.
Streptococcus sanguinis (S. sanguinis) is an oral commensal and early colonizer of the tooth surface that contributes to dental biofilm homeostasis. Zinc oxide nanoparticles (ZnO NPs) are often incorporated into dental restorative materials to enhance mechanical performance and confer antibacterial properties; however, their effects on S. sanguinis have not been thoroughly studied. Here, we investigated the antimicrobial and antibiofilm efficacy of ZnO NPs against this bacterial species. ZnO NPs exhibited a minimal inhibitory concentration (MIC) of 100 µg/mL and caused rapid, dose-dependent suppression of intracellular ATP levels and overall metabolic activity within 2-4 h of exposure. ZnO NPs induced reactive oxygen species (ROS) production in a dose-dependent manner. The free radical scavenger α-tocopherol partly prevented the antibacterial effect of ZnO NPs, suggesting that lipid peroxidation contributes to ZnO NP-mediated toxicity, although it is not the sole mechanism involved. Short-term exposure (2 h) to ZnO NPs did not significantly affect membrane integrity or cellular morphology, whereas prolonged treatment (24 h) resulted in pronounced membrane permeabilization, membrane hyperpolarization, and cellular swelling. Computational morphometric analyses of high-resolution scanning electron microscopy (HR-SEM) images of planktonic growing bacteria after a 24 h treatment confirmed a significant, dose-dependent increase in cell surface area and surface roughness. Importantly, ZnO NPs also reduced the metabolic activity and compromised the structural integrity of mature, preformed biofilms. Collectively, these findings demonstrate that ZnO NPs exert antimicrobial and antibiofilm effects against S. sanguinis through early metabolic inhibition associated with oxidative stress followed by progressive membrane dysfunction.
Objective.Visual impairments create significant challenges for navigation. This work explored the potential for an autonomous navigation aid with multisensory feedback to improve navigational performance for users of visual neuroprostheses.Approach.An autonomous navigation system was developed that maps the environment in real time and provides guidance using combinations of prosthetic vision, haptic, and auditory cues. Navigational performance was evaluated in 20 sighted participants using simulated prosthetic vision and in a single-subject case study of an Argus II visual neuroprosthesis user. Participants completed three tasks: navigate to destination, obstacle field traversal, and relative distance judgment. Multiple sensory feedback configurations incorporating visual, haptic, and auditory cues were compared. Performance metrics included collision rate, distance traveled, task completion time, navigation success rate, and accuracy of relative distance judgments.Main results.Performance differences across sensory configurations were most pronounced in navigation success and collision rates. Haptic plus audio feedback was highly effective for navigation tasks, enabling successful navigation in nearly all trials involving haptic guidance. Argus vision (AV) alone was inadequate for navigation. Depth vision (DV) provided modest improvements over AV but did not enhance performance beyond haptic and audio guidance when combined. Wide field-of-view DV yielded additional benefits, particularly for obstacle field traversal where its performance exceeded other modes. Adding AV to haptic and audio also provided no benefit and, in some cases, degraded performance. Performance trends for the Argus user were generally comparable to those of sighted participants across sensory modes, with the exception of the relative distance judgment task, in which the Argus user demonstrated better performance. Among sighted participants, increased field of view and resolution independently improved relative distance judgment accuracy.Significance.These findings demonstrate the potential of multimodal feedback systems to improve navigation for prosthetic vision users. (ClinicalTrials.gov NCT04359108).
Software agents are beginning to shop and pay on a person's behalf. Agent payment protocols such as AP2 produce cryptographically valid signatures for completed purchases, yet do not constrain the decisions that lead to them. Consequently, ordinary product-description text can steer a shopping agent into forming a cart that passes every protocol check but no longer matches the user's request. In this paper, we show that this vulnerability enables three related attacks. In the first attack, the agent is steered into fetching another user's payment credentials. In the second, it assembles a cryptographically valid cart whose contents do not match what the user was shown. In the third, a single factual claim about stock or product lineage moves the agent from the cheaper displayed item to a more expensive one, while the resulting cart remains fully consistent with the listing. In experiments using the Gemini Flash-Lite models that AP2's sample agents specify by default, the three attacks succeeded at rates of 90
Background Dialysis nurses routinely make high-stakes clinical decisions under conditions of uncertainty, balancing protocol-based guidelines with contextual and experiential judgment. Recent advances in artificial intelligence (AI) raise questions regarding its potential role in supporting nursing clinical reasoning. Aim To compare clinical reasoning performance across experienced dialysis nurses, a general-purpose large language model (ChatGPT-4), and an agent-based AI system (MAI-DxO) using real-world nephrology scenarios, and to explore patterns of human nursing decision-making. Design A comparative, scenario-based study. Methods One hundred and ten dialysis nurses and two AI systems independently responded to four validated hemodialysis scenarios reflecting common clinical dilemmas. Responses were evaluated by senior nephrology nursing experts for diagnostic accuracy, appropriateness, and quality of clinical reasoning. Results The agent-based AI system achieved higher mean scenario scores than both ChatGPT-4 and nurses, particularly in structured justification and differential diagnosis. Nurses demonstrated greater variability, with strengths in contextual interpretation and recognition of dialysis-specific complications. Cluster analysis identified three distinct nursing reasoning profiles: protocol-driven, holistic-explanatory, and minimalist. Conclusion While AI systems can provide structured and guideline-consistent clinical reasoning, experienced dialysis nurses contribute contextual judgment and practical insight that remain essential to safe patient care.These findings support a complementary, rather than substitutive, role for AI in nursing clinical decision-making.
