Sharding architectures enable linear scalability for block- chain systems but introduce a critical vulnerability: adaptive adversaries can concentrate malicious nodes within individual shards to subvert consensus. Traditional Byzantine Fault Tolerance protocols assign equal voting power to all validators, making them vulnerable when adversaries achieve numerical superiority within a shard. Existing approaches rely on probabilistic defenses through random validator assignment, which provide no deterministic guarantees against patient adversaries. This paper introduces DynaGuard, a trust-weighted consensus mechanism that integrates behavioral reputation directly into BFT protocols for permissionless sharded networks. DynaGuard synthesizes economic stake, network longevity, and historical behavior into a dynamic Trust Score that determines voting power. Consensus requires two-thirds cumulative trust weight rather than node count, significantly altering adversarial calculus by requiring sustained honest behavior to accumulate influence. We establish formal safety and liveness properties under adaptive adversaries controlling less than one-third of network trust. Comprehensive simulation across varying network scales demonstrates that DynaGuard reduces shard failure rates by 98.6% compared to standard PBFT under 33% adaptive adversaries, while introducing 3.2% communication overhead and 5.8% latency increase, confirming practical viability.