Security patches, AI SRE models, and agent-aware tracing

Security and safeguards under the tracing hood

The week's most serious thread ran through Datadog's tracing libraries. A security vulnerability in tracing, described in GHSA-3633-r77q-h3v3, got patched in dd-trace-py's stable line, which also closed a quiet evaluation gap in AI Guard: LangChain model responses were never checked for unsafe output, so they were neither reported nor blocked. The release candidate that preceded it stopped AI Guard credentials from leaking into debug logs, exception messages, and span error tags along the way. Both stabilized and extended the Python SDK with tracing stats cardinality limits, Trio trace-context support, Azure Durable Functions tracing, and LLM Observability sampling rules. On the JavaScript side, dd-trace-js picked up the same Test Optimization work on its 6.x line — WebdriverIO failure videos and Cypress suite TIA skip counts — plus per-trace snapshot sampling for Dynamic Instrumentation and a fix so LLM Observability events reach the Lambda extension before the sandbox freezes; the 5.x maintenance line carries the identical changes for teams still on the older major.

Elsewhere in Datadog's stack, the Agent locked down Private Action Runner Kubernetes access by default, requiring an explicit allowlist of custom resources, and stopped overwriting customized script configuration on upgrades — plus a small but welcome fix for NVML's packed bit-error-rate encoding, which had been reporting 4095 for perfectly healthy GPU links. The Browser SDK turned its attention to WebSockets, collecting vitals for them and firing a point-in-time RUM action on the first hit of browser and EXIT probes.

Datadog's AI SRE gets its own model

The headline launch was Scout 1.0, Datadog's first AI model trained specifically for investigating production issues, now powering Bits Investigation. Trained with reinforcement learning over historical incidents and telemetry simulations, Scout is a cost play as much as an accuracy one: Datadog claims its medium mode matches top-tier general models at a fraction of the inference cost, making it cheap enough to run investigations far earlier in an incident. Two follow-on posts rounded out the week: RUM Operations Monitoring, free with RUM, lets teams instrument application-controlled steps like "process payment" with their own availability and latency SLOs, and Datadog shared how it judges service criticality across roughly 6,000 internal services by blending postmortem history, error traces, traffic, and on-call data. For the infrastructure-minded, there's a walkthrough of how Kubernetes DRA lifts GPU utilization past the limits of the legacy device plugin model.

Sentry's plugins, mobile, and Python surface all shift

Sentry's biggest structural change was releasing its bundler plugins from the SDK repo, which jumps their version from 5.4.1 to the SDK's and raises the floor to Node 20.19. New projects should import the re-exported @sentry/bundler-plugins entry point. The surrounding releases added Cloudflare cron monitoring and Neon DB spans, and quieted a duplicate process session that Node servers were sending once request sessions were tracked. Mobile got attention too: the React Native SDK records Turbo Module calls, capturing which native module and method were live at a native crash, and the Android SDK began experimental Jetpack Navigation 3 support via SentryNavEffect. On the server side, Sentry Python deprecated send_default_pii and transaction profiling and added localhost and user-agent attributes for inbound filters. In the product itself, Explore > Agents can now sort and filter conversations by whole-conversation cost, tokens, tool calls, and errors.

Labels and a quieter Axiom release

Axiom shipped labels for organizing resources: assign colored tags like env=production or team=platform to datasets, dashboards, monitors, and notifiers, filter lists by ANY or ALL matches, and let the MCP server surface that context to agents so they know which service, environment, and owner a resource belongs to.