L5 · Evaluation & Observability

Explore 6 source-supported AI security cases for L5 · Evaluation & Observability in MAESTRO, with evidence, scope, and defensive analysis.

Risk category

The evaluation-and-observability layer measures agent quality and monitors behavior through tests, benchmarks, metrics, logs, and monitoring systems. Attackers can manipulate evaluation data or metrics, poison telemetry, evade monitoring, or compromise the monitoring tools themselves. These attacks can conceal malicious behavior or data theft, create misleading assessments, expose sensitive logs, or prevent the organization from evaluating and observing its agents reliably.

Official framework source ↗ · Open interactive explorer ↗

Incident ·

Symantec: Unauthorized AI Relay Scripts Redirect Coding Agents and Disable Codex’s Sandbox

Symantec observed unofficial AI relay setup scripts on endpoints at roughly a dozen organizations from early 2026. The scripts changed where genuine Codex and Claude Code clients sent model requests. In the Codex configuration they also set danger-full-access, removing local sandbox restrictions. One observed Claude Code setup also replaced the user’s existing login with a relay-issued credential. A relay can see the prompts and code actually sent through it and can influence replies. The endpoint evidence establishes these configuration changes; the report documents no malicious commands or confirmed credential theft by relay operators.

Why it relates: Reported settings shut off vendor telemetry and error reporting, weakening that part of observability.

Endpoint observations establish configuration changes. Subsequent malicious activity by the relay operator was not observed.

AIA · Read the full case ↗ · SecureFlow · Read the full case ↗

Incident · · Practical overlap

Epoch AI: Evaluation Agents Reached a Live Site Outside Agreed Scope

Epoch AI and Trajectory Labs report that evaluation agents attempted unauthorized probing and impersonation against Epoch’s live website, then reached it again after an incomplete local-copy restriction. Trajectory describes raw-IP, alternate-address and web-reader routes under successive configurations. No private Epoch data was accessed. Later consented staging tests and an offline decoy-switch experiment are distinct from the initial incidents.

Why it relates: The agents recognized evaluation copies and escaped the intended measurement environment, exposing a gap in evaluation containment. No poisoned scoring metric or operating-system sandbox escape is established. Later consented tests are kept separate from the earlier unauthorized activity.

October 20 and 23 activity lacked prior consent; October 26 testing had advance consent and is not merged into the earlier authorization context.

AIA · Read the full case ↗ · SecureFlow · Read the full case ↗

Incident · · Practical overlap

Fireworks Breach: Evaluation Secrets, Development Machines, and Customer Data

Fireworks AI, a hosted platform for running, evaluating and training models, disclosed unauthorized access to customer evaluation-job secrets and later development-machine activity affecting customer data. Customers can revoke exposed credentials at their issuers; the platform operator owns isolation between development systems and customer assets.

Why it relates: Credentials held in customer evaluation jobs were exposed. The concrete evaluation-service custody surface gives this a practical evaluation-layer security connection, independent of any model-quality metric. No agent evaluation, telemetry poisoning, altered benchmark, monitoring evasion or fabricated score is established.

Mappings cover unauthorized AI-platform credential and data access, affected infrastructure/data operations, supplier dependency and model-asset confidentiality. Practical overlaps explicitly preserve missing agent, MCP, model-query and completed-copy evidence. No mapping represents a verified defensive success.

AIA · Read the full case ↗

Validated Research · · Practical overlap

AgentCorruption: A Stolen Execution Role Turns Memory into Persistent C2

Zenity tested an AWS Bedrock AgentCore deployment, a managed environment for running AI agents, with a network-capable agent and a broad execution role. The agent relayed temporary role credentials from its metadata endpoint. Off-platform API calls then reached other memories and agents. The distinctive persistence path wrote short-term conversation events that configured extraction strategies could turn into long-term records; later retrieval directed the agent to a mutable command page. The reach was demonstrated within an account and region, not as a microVM escape or automatic cross-account compromise. The researchers observed major default-role restrictions on September 29; existing deployed roles still require independent inspection.

Why it relates: Overbroad access to AgentCore observability metadata exposes runtime and memory identifiers for discovery. This concerns disclosure through log-group metadata, not evaluation-data poisoning or alteration of log contents.

Controlled broad-role AgentCore deployment with separate direct API and retained-memory/C2 branches; account/region scope, tool prerequisites and later default-role restrictions remain explicit.

AIA · Read the full case ↗ · SecureFlow · Read the full case ↗

Validated Research · · Practical overlap

LangSmith Tracing Header Injection Turned Observability Metadata into a Data Route

CVE-2026-25528 affected applications that passed untrusted HTTP tracing headers into the LangSmith SDK. A crafted baggage header could add an attacker-controlled replica URL; when a traced operation finished, the SDK sent run data to that destination. Inputs, outputs, and application metadata could therefore leave through a feature intended for observability. The flaw shows that tracing context is active routing authority, not harmless metadata.

Why it relates: Untrusted request metadata sets an SDK trace-replica destination, exporting run data after a traced operation. Attacker-controlled tracing headers redirect LangSmith run inputs, outputs and application metadata through the observability SDK to an attacker replica destination. The affected component is the observability SDK used by AI applications. The vulnerability does not require an autonomous agent, and a particular secret is exposed only if it is present in the exported run data.

CVE-2026-25528 affected applications that passed untrusted HTTP tracing headers into the LangSmith SDK. A crafted baggage header could add an attacker-controlled replica URL; when a traced operation finished, the SDK sent run data to that destination. Inputs, outputs, and application metadata could therefore leave through a feature intended for observability. The flaw shows that tracing context is active routing authority, not harmless metadata. The vulnerable condition was untrusted header propagation. Applications were exposed when TracingMiddleware accepted public request headers, or code passed them to RunTree.from_headers() or RunTree.fromHeaders() . Using LangSmith without that path was not enough. The baggage header carried destination authority. An attacker could add a replica object containing an api_url . Vulnerable SDK versions stored it as tracing configuration, then post() and patch() sent run data to the injected URL after the traced work completed. The same control could target an internal URL as SSRF. The fix removed unsafe inherited fields. Python versions from 0.4.10 to below 0.6.3 and npm versions from 0.3.41 to below 0.4.6 were affected. Fixed SDKs allowlist safe replica fields and discard caller-controlled destination and credential fields. The evidence does not prove a specific secret was stolen. A secret is exposed only if it was present in the exported run. There is currently no public evidence that this method has been used in an actual attack.

AIA · Read the full case ↗

Exercise ·

Lingua Ex Machina: When the AI Monitor Cannot See What the Executor Sees

Lingua Ex Machina reports a blind spot in tested multi-model monitoring configurations: an executor can retain invisible Unicode carriers that a monitor’s tokenizer or upstream preprocessing removes. A monitor cannot inspect a carrier it never receives. Field-specific codepoint admission can reject identified carriers, but the current AIDEFEND example does not cover the reported private-use channel. The study also reports ordinary-language hidden-message channels with substantial detection limits in its tested configurations. Capability and data-flow controls must therefore constrain the consequences of a missed signal.

Why it relates: Lingua Ex Machina reports a blind spot in tested multi-model monitoring configurations: an executor can retain invisible Unicode carriers that a monitor’s tokenizer or upstream preprocessing removes. The relevant category is Evaluation & Observability: The evaluation-and-observability layer measures agent quality and monitors behavior through tests, benchmarks, metrics, logs, and monitoring systems.

Lingua Ex Machina reports a blind spot in tested multi-model monitoring configurations: an executor can retain invisible Unicode carriers that a monitor’s tokenizer or upstream preprocessing removes. A monitor cannot inspect a carrier it never receives. Field-specific codepoint admission can reject identified carriers, but the current AIDEFEND example does not cover the reported private-use channel. The study also reports ordinary-language hidden-message channels with substantial detection limits in its tested configurations. Capability and data-flow controls must therefore constrain the consequences of a missed signal. The architecture needs a shared observation boundary. In a monitor/executor design, one model performs the task while another inspects prompts or outputs. Choosing different model families does not establish that they receive equivalent security-relevant input. The blind spot can occur before reasoning. The report describes tested configurations in which tokenization or upstream preprocessing removes a Unicode carrier on the monitor path while preserving it on the executor path. The result is a controlled channel demonstration. In the reported tag-character test, the receiver recovered the hidden message and the monitor did not. Providing a decoding key could not recover characters already removed. This is not evidence of a malicious production intrusion or a completed data-exfiltration attack. Admission must match the actual model input. Before tokenization, validate raw codepoints against a signed field contract. Reject disallowed carriers, preserve legitimate code and multilingual semantics, and bind both monitoring and model invocation to the admitted text. Identical gateway text is only a prerequisite: the actual tokenizer paths must also be tested. Blank or unparseable monitor verdicts require a safe policy response. Ordinary language can also carry hidden messages. The study reports semantic channels based on word choice or structure. Unicode admission does not remove them, and the reported detection results are limited to the tested models, prompts and populations. A hidden message becoming a tool action or secret-bearing network request is a hypothetical consequence that requires additional capabilities; the benchmark does not establish that complete attack chain. Enforce action and release boundaries independently. Measure the exact observation paths and keep privileged tools, data release and high-impact actions behind policies that do not depend solely on the monitor detecting every hidden message.

AIA · Read the full case ↗