Aporia

What is Aporia?

Aporia is an observability and Guardrails platform designed for developing AI applications; it used to serve both traditional machine learning and generative AI. It implements security policies between the inputs and outputs of models, and it keeps records of sessions, violations, and model performance.

Aporia was acquired by Coralogix in December 2024, and its team as well as its technology were integrated into Coralogix AI. For current purchasing tasks and the acquisition of new capabilities, Coralogix AI Center should be used as the main platform.

Current product status

• The Aporia brand’s website still contains product information and historical documents.
• The original Guardrails and observability technologies have been integrated into Coralogix.
• The Coralogix AI Center provides a unified view of AI performance, quality, security, and governance.
• The former head of the Aporia team continues to be involved in Coralogix’s AI products.
• New customers should confirm the migration, support, and contracting entity.
• Old tutorials and old prices cannot be directly considered as the current options.

What is AI Guardrails?

Guardrails are independent verification layers that operate before user input is fed into the large model, as well as before the model’s results are returned to the user. They can allow, block, replace, or modify content, thereby eliminating the need to include all security rules within the system’s prompts.

Primary protection capabilities

• Detection indicates prompt injection and jailbreak attempts.
• Identify and process personal identification information.
• Check for toxicity, hatred, violence, and inappropriate content.
• Determine whether the response falls outside the allowed or restricted topics.
• It assists in detecting hallucinations that lack contextual support.
• Use custom rules to enforce corporate policies.
• Different policies are configured for input and output respectively.

Prompt injection protection

Prompt injection strategies are used to detect inputs that attempt to override system commands, steal context, or force tools to act beyond their authorized scope. They can reduce risks, but they cannot guarantee the prevention of all new types of attacks.

Hallucinations and correctness checking

The platform can assess whether a response is well-supported based on the provided context, or it can use evaluation strategies to flag suspicious results. Open-world facts, professional conclusions, and real-time information still require reliable data sources and human review.

Protection of sensitive information

• Detect personal information in user input.
• Prevent the model from outputting data that should not be disclosed.
• Mask or rewrite specific fields.
• Record the reasons for the policy trigger and the actions taken.
• Differentiate data protection rules by application.
• Integrated with the company’s existing privacy processes.

Content security policy

• Restrict content involving violence, sex, hate, and harassment.
• Set allowed and prohibited topics according to the business requirements.
• Convert high-risk responses into safe replies.
• Block obviously malicious requests at the input stage.
• Intercept content that causes the model to lose control at the output stage.
• Train custom business boundaries through examples.

Custom Guardrails

Teams can create custom rules in accordance with business policies, and use positive and negative examples to indicate what should be allowed or blocked. Custom policies are suitable for industry-specific terminology, brand guidelines, and internal processes; however, it is necessary to continuously test for false positives and false negatives.

Multimodal security

Aporia has publicly supported Guardrails for AI applications involving text, images, and audio. The specific input formats, models, and policies that are currently supported should be verified within the Coralogix product environment.

AI application observability

• Track prompt tokens, responses, and tool calls.
• Check latency, errors, tokens, and usage.
• Analyze the quality changes across different models and versions.
• Troubleshoot issues by application, session, and user path.
• Identify the policy trigger rate and risk type.
• Link AI execution with traditional application telemetry.

Session Explorer

The session browsing feature allows users to view consecutive conversations, the decisions made based on certain policies, and the final outcomes. Teams can identify issues by examining individual user sessions, rather than relying only on summary charts.

Traditional machine learning monitoring

• Monitoring data drift and concept drift.
• Missing values, outliers, and data integrity issues were detected.
• Track how model performance changes over time.
• Compare the performance across clusters and business dimensions.
• Use explanatory capability to investigate anomaly prediction.
• Trigger an alert or automate a workflow.

Which users are it suitable for

• Engineering teams preparing to put generative AI into production.
• Companies that need to monitor the quality of RAG-based question-answering.
• AI products for processing customer data and sensitive content.
• Organizations need to standardize AI and application observability.
• Platform team for managing multiple models and agents.
• Data teams that need traditional machine learning drift monitoring.

Typical use cases

• To prevent the leakage of sensitive information by customer service robots.
• Detect prompt injection in enterprise knowledge assistants.
• Restrict financial or medical assistants from answering questions that fall outside their scope.
• Track tool calls and reasons for failures of AI agents.
• Evaluate whether the RAG response is supported by the context.
• Monitor data drift in credit, recommendation, and risk control models.
• Link AI anomalies to application logs and infrastructure issues.

It’s not very suitable for which situations

• Users who only engage in private conversations and have no need for production monitoring.
• It is hoped that Guardrails will replace all manual review processes in high-risk operations.
• Teams that need a complete, open-source security engine for offline deployment.
• Early prototypes lacked unified logging and application identifiers.
• Customers who only wish to purchase the old Aporia standalone package.
• Teams lacking the necessary manpower to carry out continuous strategy testing and operations.

Tutorial on Connecting to Guardrails

1. List the data types that AI applications receive and generate.
2. Identify the risks that must be prevented, altered, recorded, or handed over to human personnel.
3. Integrate the current Guardrails SDK or proxy into the testing project.
4. First, enable a small number of preset policies and use historical samples for playback.
5. Calculate the false positives, false negatives, and delays for each strategy.
6. Add examples of approvals and denials for business rules.
7. Activate blocking only after verification in observation mode.
8. After going live, continuous monitoring is carried out to track session and policy trigger trends.

Tutorial on configuring prompt injection protection

1. Collect samples of known jailbreaks, instruction overwrite, and data theft.
2. Distinguish between normal user commands and actual attack attempts.
3. Enable the prompt injection strategy at the input end.
4. Set up logging or manual handling for events with low confidence.
5. Configure blocking and secure responses for attacks with high confidence.
6. Test multilingual, encoding, and indirect injection variants.
7. Confirm that blocking will not disrupt normal business requests.
8. Regularly expand the test set based on attack logs.

Create AI monitoring tutorials

1. Establish separate project and environment identifiers for each AI application.
2. Collect prompt words, responses, latency, tokens, and error messages.
3. First, apply the necessary masking to sensitive fields.
4. Define quality, safety, cost, and performance metrics.
5. Set the application owner and the person who will receive alerts.
6. Link abnormal sessions to application logs and release versions.
7. Establish a process for reviewing issues and updating strategies.
8. Evaluate the protection effectiveness and AI Unit consumption on a monthly basis.

Current billing method

Following the acquisition of Aporia, Coralogix’s AI assessment services and Guardrails utilize AI Units for billing. AI Units are calculated separately from the quotas related to logs, metrics, and traceability; assessment and protection services typically consume resources based on the number of tokens processed and the type of functions used.

Price and version details

How do AI Units understand things?

AI Units are separate consumption pools created by Coralogix for AI-related functions; they cannot be exchanged for log, metric, or Trace units. The transmission of AI pathway data is billed according to the standard Trace pricing scheme, with additional costs associated with AI functions only arising when evaluation or Guardrails are enabled.

What needs to be confirmed before making a purchase?

• Methods for migrating existing Aporia accounts and preserving data.
• Models, languages, and input types supported by Guardrails.
• Free AI quota, overage pricing, and monthly budget control.
• The combined cost of logs, traces, and AI Units.
• Data regions, privacy agreements, and security certifications.
• Support levels, SLAs, and incident response processes.
• Plan for migrating from the old SDK to Coralogix AI Center.

False positives and false negatives

• Excessively strict policies may prevent normal user requests.
• A too-permissive strategy might fail to detect new types of attacks.
• Multilingual terms and industry jargon can affect the accuracy of classification.
• Long contexts may increase latency and make judgment more difficult.
• Hallucination detection relies on the available reference context.
• The effectiveness of a strategy must be evaluated using real business examples.

Performance and latency

Guardrails introduce additional processing steps in the model pipeline; therefore, it is necessary to measure the average value, tail latency, and the behavior of the system in case of failures. For critical applications, it is also important to decide whether to reject requests or to allow them to pass temporarily when the protection services are unavailable.

Privacy and data processing

• The prompts and responses may contain customer and employee information.
• The minimum necessary fields should be determined before starting a collection session.
• Pre-mask passwords, tokens, and payment information.
• Limit who can view the full session.
• Set up data retention, deletion, and audit processes.
• For cross-border or regulated data, it is necessary to verify the contract and the region.

Security boundary

Guardrails are part of the layered defense strategy; they cannot replace identity authentication, access control, tool authorization, network isolation, and manual approval. Even if a content policy is approved, an agent may still carry out dangerous actions due to incorrect permissions.

API and integration capabilities

Originally, Aporia allowed access to generative AI applications through SDKs or proxies, and it could be integrated with OpenAI’s call chain. Currently, Coralogix AI Center handles monitoring and protection via AI connections, SDKs, and platform policies; the specific interfaces are subject to the current documentation.

GitHub and the open-source status

The official Aporia GitHub repository contains import tools, sample projects, and MLOps resources, but the core Guardrails platform is not a fully open-source product. The licenses for the public SDKs and samples do not extend automatically to the hosted detection engines.

Product advantages

• Real-time protection is implemented on both the input and output sides of the model.
• It covers prompt injection, sensitive information, and content security.
• Predefined policies and enterprise-customized rules are supported.
• It provides session-level investigation and policy-triggered analysis.
• It provides monitoring for both generative AI and traditional machine learning.
• After being integrated with Coralogix, it enables the association of application and infrastructure telemetry.

Usage restrictions

• Aporia has been acquired, and the product access points as well as the contracts have changed.
• Security detection cannot achieve zero false positives and zero false negatives.
• This strategy will increase latency and the cost of additional AI Units.
• The assessment of hallucinations depends on the context and task definition.
• High-risk decisions still require human supervision.
• The core testing and hosting platform is not open source.
• Old documents and old packages may no longer be applicable to new customers.

Basic information

Recommendation score

The comprehensive recommendation score is 4.2 out of 5 points. Aporia’s Guardrails technology is suitable for ensuring AI security and monitoring, but new users should evaluate Coralogix AI Center first, taking into account factors such as the accuracy of policies, latency, and cost of usage during the pilot phase.

Frequently Asked Questions

Can Aporia still be used?

The related technologies are still in use, but they are currently primarily integrated into the Coralogix AI Center.

Who acquired Aporia?

In December 2024, Coralogix announced the acquisition of Aporia.

What can Aporia Guardrails prevent?

It can assist in detecting risks such as prompt injection, sensitive information, toxic content, off-topic content, and hallucinations.

Can guardrails ensure AI safety?

No, it must be used in conjunction with access control, monitoring, testing, and manual review.

How is the pricing done now?

In Coralogix, evaluation and Guardrails are primarily billed using separate AI Units.

Is there a free quota?

The free plan includes a monthly AI quota, which is reset and subject to the account rules.

Is traditional machine learning supported?

Throughout its history, Aporia has provided monitoring for drift, data quality, performance, and interpretability.

Is Aporia open source?

The core platform is not open-source; the official GitHub page only makes available some tools, examples, and resources.

Why is manual review still needed?

Tests can yield false positives and false negatives; for high-risk operations, the ultimate responsibility cannot be entrusted to a single automated rule.