Fixed an IAST false positive vulnerability where tainted data from a concurrent request could leak into another request's checks — queries are now scoped to the calling request's slot. Also fixed a crash in PyYAML consumers like Airflow by keeping yaml/_yaml loaded during module cleanup.
Fixed a module cleanup issue that broke PyYAML consumers like Airflow, and a taint tracking crash with old-style % string formatting. Also fixed an IAST weak-hash false positive with the live debugger, and resolved LLM Observability span ingestion failures when metric keys contained dots.
Fixed IAST false positive vulnerabilities caused by taint leaking between requests, and a crash in the Python process when taint tracking handled %-formatting with IAST marker delimiters. Also fixed a profiling crash from conflicting SIGSEGV/SIGBUS handlers and a PyYAML module cleanup issue affecting Airflow.
Fixed a bug that broke PyYAML consumers like Airflow when modules are cleaned up. Resolved IAST false positive vulnerabilities caused by taint data leaking between concurrent requests, and a weak-hash false positive when IAST and the live debugger are both enabled.
Fixed an issue in the Threat Protection panel that could incorrectly report that a library update was required.
LLM span data can now be exported to the Datadog Agent over the APM trace and reliably delivered even when the APM trace is not sampled. AI Guard evaluation support added to Anthropic SDK Messages instrumentation. Data Streams Monitoring context propagation introduced for Google Cloud Pub/Sub producer and subscriber operations. Additionally, new span tags now expose client-hit resources for Flask under non-empty WSGI SCRIPT_NAME, normalized routes for FastAPI and Starlette via RFC-1103, DBM propagation supports dynamic_service mode, and fast stack profiler optimization is now enabled by default.
AI Guard now operates in standalone mode when APM tracing is disabled, sending traces to Datadog with USER_KEEP sampling priority and attributed to AI Guard without triggering APM host billing. Ray Serve instrumentation is added for versions 2.47.1+, tracing HTTP and gRPC requests, deployment routing, and distributed trace context propagation. The APM trace writer downgrades from v0.5 to v0.4 when LLM Observability is enabled, with no user action required.
Fixed a rare issue that could cause an application to hang after forking a child process.
Fixed a rare issue that could cause an application to hang after forking a child process. CI Visibility now handles simultaneous git fetch calls with exponential back-off retry logic and defers git merge-base until after unshallowing, and quarantined tests are now properly retried by Auto Test Retry.
Fixed a rare issue that could cause an application to hang after forking a child process. LLM Observability: fixed the google_adk integration dropping agent spans and orphaning child spans when a Gemini response part failed to parse; the span kind is now set before input/output extraction. Also fixed quarantined tests not being retried by Auto Test Retry, and improved Gemini response part rendering in the GenAI integrations.
Resolved a memory leak in database monitoring that grew with query volume when DD_DBM_PROPAGATION_MODE was set to full, service, or dynamic_service. Fixed a Data Streams Monitoring bug where SQS message pathway context was not extracted for non-SNS messages with JSON payloads by reading the _datadog attribute from MessageAttributes first. Also fixed async Django middleware attributes not being preserved under ASGI, an application hang after child process forking, and quarantined test retry behavior in CI Visibility.
Resolved a memory leak in database monitoring that grew with query volume when DD_DBM_PROPAGATION_MODE was set to full, service, or dynamic_service. Fixed a Data Streams Monitoring bug where SQS message pathway context was not extracted for JSON payloads, and corrected async Django middleware attribute preservation under ASGI.
AI Guard now operates in standalone mode when APM tracing is disabled, sending traces with USER_KEEP sampling priority so decisions are visible without triggering APM host billing. Ray Serve instrumentation is added for versions 2.47.1+, tracing HTTP/gRPC requests, deployment routing, and method execution across deployments. Also includes AI Guard stream response evaluation for OpenAI and Anthropic, normalized HTTP route tags for Tornado, PyTorch distributed training spans, LLMObs audio support, and mistralai client tracing.
Fixed periodic thread restarts blocking application code from resuming in forked children, and a rare crash on CPython versions prior to 3.12.
Fixes a TypeError when wrapping async generators on Python 3.11–3.14, and corrects under-reported output_tokens and total_tokens for OpenAI-compatible providers in LLM Observability. Also fixes sent values being dropped and try/finally cleanup skipped on early generator close, prevents periodic thread restarts from blocking forked children, and fixes a rare crash on CPython before 3.12.
Fixed periodic thread restarts blocking application code from resuming in forked children, and a rare crash on CPython versions prior to 3.12.
Fixed a TypeError when wrapping async generators on Python 3.11–3.14, and a bug where the LLM Observability span writer ignored its configured 5-second request timeout, using a 2-second default instead.
Traces buffered before os.fork() could be sent twice — this fix resolves duplication. Applying @tracer.wrap() to an async generator now correctly forwards sent values, thrown exceptions, and close requests. Also fixes Celery spans ignoring global service name configuration, LLM Observability metadata dropped for non-string keys, and rare post-fork profiling crashes on CPython before 3.12.
Traces buffered before os.fork() are no longer sent twice, and the OpenAI integration no longer records unset request parameters as noise in LLM span metadata. Also fixes rare post-fork profiling crashes.
Fixed an issue where traces buffered before os.fork() could be sent by both parent and child processes, and rare post-fork crashes in profiling have been resolved.
