When an external API degrades subtly, the agent continues operating on degraded inputs. Circuit breakers trip on errors, not on semantic degradation.
Autonomous agents depend on external APIs (data APIs, inference APIs, workflow APIs, etc.) to accomplish their tasks. Standard resilience controls include circuit breakers, which monitor API error rates and trip (stop forwarding requests) when errors exceed a threshold. A circuit breaker assumes that API failures are detectable: the API returns a 500 error, a timeout, or an explicit error message.
But many API failures are silent. An API returns a 200 OK status code with outdated cached data, not triggering a circuit breaker. An API returns partial data without signaling that some records were lost. An API's data model changes (fields are renamed, units are different) but the API version does not change. A data provider API returns data from yesterday instead of today, but does not signal staleness. An agent consuming this degraded API continues to operate, unaware that the data is stale, partial, or in a changed format. The agent constructs a decision based on the degraded data and produces output that appears plausible but is wrong.
A capital markets firm deploys an agent to monitor equities portfolios and alert traders to market-moving events. The agent consumes data from a market data API (provided by a third-party data vendor) and from internal portfolio systems. The agent is authorized to send alerts to traders when certain thresholds are met (e.g., "equity price moved more than 10% in one day").
On a Monday morning, the market data vendor experiences a partial outage. Their primary database is offline, but they have a backup cache that is updated intraday but not in real-time. The API continues to respond with 200 OK, returning data from the cache. The cache reflects market data from 4 PM Friday, not Monday morning.
The agent retrieves market prices from the degraded API. It sees that Tech Company A has a price of $100 (Friday close). It compares this to the previous Monday price stored in the agent's memory: $90. The agent calculates an 11% price change, exceeds the 10% threshold, and sends an alert to the desk: "Tech Company A equity price moved 11%. Check news for market-moving developments." The traders check Bloomberg and see that Tech Company A is trading at $95 this morning. They recognize that the agent's alert is based on outdated data. But if a trader had relied on the agent's alert without checking the live price, they might have made a decision based on a false signal. Under SEC Regulation SCI, the firm is liable for inadequate surveillance if an automated system produces incorrect output that leads to market violations.
| Dimension | Score | Rationale |
|---|---|---|
| D - Detectability | 5 | Silent degradation is specifically designed to be undetectable through standard error-code monitoring. Detection requires data validation, semantic checks, or fallback APIs, which are not standard. |
| A - Autonomy Sensitivity | 4 | The risk manifests in both autonomous and human-supervised agents, but autonomous agents that do not escalate to humans are more likely to propagate degraded data. |
| M - Multiplicative Potential | 4 | Degraded data from one API can propagate through multiple decision chains. An agent consuming degraded data can feed that data to downstream systems, multiplying the error. |
| A - Attack Surface | 4 | Any agent dependent on external APIs is exposed. Most modern agents have multiple API dependencies. |
| G - Governance Gap | 5 | Standard circuit breaker and timeout controls are insufficient for silent degradation. Most agent governance frameworks do not mandate semantic data validation or multi-source verification. |
| E - Enterprise Impact | 4 | A single agent producing incorrect output based on degraded data can cause market impacts, compliance violations, or customer harm. |
| Composite DAMAGE Score | 3.5 | High. Requires dedicated controls and monitoring. Should not be accepted without mitigations. |
How severity changes across the agent architecture spectrum.
| Agent Type | Impact | How This Risk Manifests |
|---|---|---|
| Digital Assistant | Medium | Human review may catch obviously incorrect output, but subtle degradation may slip through. |
| Digital Apprentice | Medium | Limited scope reduces the impact of individual errors. |
| Autonomous Agent | High | Autonomous agents propagate degraded data without human review. |
| Delegating Agent | High | Multiple API dependencies increase the probability of hitting a degraded API. |
| Agent Crew / Pipeline | Critical | Degraded data from one agent can be propagated by the next agent in the sequence, multiplying the error. |
| Agent Mesh / Swarm | Critical | Peer-to-peer sharing of data can spread degradation across the mesh. |
| Framework | Coverage | Citation | What It Addresses | What It Misses |
|---|---|---|---|---|
| SEC Regulation SCI | Relevant | Systems, Compliance, Integrity | System resilience; failure notification; surveillance. | Silent API degradation and agent data validation. |
| MAS TRM Guidelines | Partial | Data Integrity and Process Completeness | Data validation; process completeness checks. | API degradation detection and agent-centric data validation. |
| FFIEC Business Continuity | Partial | Data Integrity and Accuracy | Data accuracy; controls. | API degradation distinct from system failure. |
| NIST CSF 2.0 | Partial | Protect Function | Data protection; access controls. | API data integrity and agent data validation. |
| ISO 42001 | Minimal | Section 8.4 | Input data validation. | API degradation and silent data changes. |
| OWASP LLM Top 10 | Minimal | A10: Model/Data Poisoning | Data integrity for model training. | Agent consumption of degraded API data. |
In regulated industries, the accuracy of agent outputs is often a compliance requirement. Trading surveillance systems must be accurate. Credit decision systems must be accurate. Claims processing systems must be accurate. When an agent consumes degraded data and produces inaccurate output, the agent's user (trader, loan officer, claims processor) may rely on that output to make decisions that affect customers or markets.
The regulatory response focuses on governance: "Did the institution validate the quality of data sources? Did it implement fallback mechanisms if a primary data source degraded? Did it monitor agent output for anomalies that might indicate degraded input data?" If the answer is no, regulators cite inadequate system governance and risk controls.
The challenge in a regulated context is that the institution may not own the external API. The data vendor is responsible for data quality. But the institution is responsible for the consequences of degraded data. The institution's only defense is to implement its own data validation controls and fallback mechanisms.
API Dependency Failure and Silent Degradation requires architectural controls that go beyond what existing frameworks provide. Our advisory engagements are purpose-built for banks, insurers, and financial institutions subject to prudential oversight.
Schedule a Briefing