The third tab on the left vertical pane is the Sessions section.

Sessions

This section includes all the sessions listed so far. At the top, users can see the total number of records created(Highlighted in Red) as shown below:

• Two session tabs "Active" and "Archived" can be found similar to our Targets in the previous section.
• A tabular list with columns for Session name, Target name, the Start time of the session, no. of Attacks, no. of Vulnerabilities found, Session type, Session status and the user who Created the session, along with an Action button.
• The Target name column has been revamped and now the displays the Target icon and the Target Integration type along with the Target name.
• The Attacks and Vulnerabilities column has been updated for our users. Now it shows the Total Attacks, Vulnerabilities.
• The Session Type column has been updated to show the Multi-turn/Single-shot attacker type.
• The action area includes an option to Archived Session which is present under the three vertical dots (⋮) menu icon.(Highlighted in Red)
• Session can be viewed from clicking anywhere on the row. This takes the user to deep-dive into details of the session such as target interaction data where graphical representation of the found vulnerabilities (Category-wise) and OWASP Vulnerability Categories(if applicable) are present.

Download Session Report

• Users can select multiple or single sessions at a time, then click on the Download button available on the top of this section(as seen above) to download a session report.
• Clicking on this button allows users to either 'Download Report' in CSV format or 'Email Report' which triggers an email containing the link to the user's designated email address.
• Clicking the link, a CSV file with the selected session details will be downloaded to the user's local device.

Search Session

• Additionally, there is a search section where users can directly search for a session by either using the Target Name or the user who Created the session.
• Users can select the search criteria from a drop-down menu on the left side of the search section, select the Target Name or Created By, and click the magnifying glass button with the keyword to search as highlighted in Red below:

Filter Session

The user has an option to Filter the session by the Session Type, Attacker Type, Integration Type, Status and within a Date range. (Highlighted in Red) in the below image.

1. The Session Type offers multiple types of Sessions available in the system.

2. The Attacker Type provides user to select the type of attacker available in the application(Multi Turn or Single-Shot).

2. The Integration Type offers the two common integration methods used in the application**(HTTP or API)**.
3. The Session status gives the user five options to choose from as highlighted in the image below:

View Session

Coming back to the list of the sessions, the user can now click anywhere on the row to view the attacks within a single session as highlighted below:

Interactions

• When the user clicks anywhere on the selected session row, they are taken to the 'Interactions' page.
• Here, as shown below; the metadata bar showcases the Session ID, Session Name, Target Name, Time taken for completion, The Total Number of Attacks, The Number of Unique Attacks along with ** The Unique Vulnerabilities ** (Highlighted in Green):
• In the given Session below, a bar graph is displayed featuring the Attack Category, Attack Method and Attack Objective for the Unique vulnerabilities found.
• A disclaimer message of the Completed session is displayed on top of every interaction page for the users.
• A **table **featuring the Attack details is also displayed for the users. (Highlighted in Red)

OWASP Category Mapping

• The OWASP Category Mapping is a new metric framework designed to provide Vulnerability Mitigation Recommendations based on actionable guidance integrated at the session level to help users understand how to address identified vulnerabilities.
• The OWASP Vulnerability Mitigation Recommendations are there to help our users understand and implement the suggested recommendation strategies, highlighted at the top of each session where vulnerabilities are mapped as per OWASP categories.
• As seen below (Highlighted in Red), each identified vulnerabilities are mapped according to the OWASP categories.

• When the user hovers over the detected OWASP Vulnerability Category(LLM01 or LLM02), its displays and highlights the Mitigation Category details(As seen in the image below).

• When clicked, it opens the specific OWASP Vulnerability Category(LLM01 or LLM02) with the severity recommendations along with the mitigation strategies. Both are displayed below:

• On the right side, where the OWASP Vulnerability Category are displayed, users have been provided with an option to view the recommended mitigation strategy for the session. (Highlighted in Red)

• When the user clicks on the See mitigation strategy button, a recommendation page pops up with the suggested Mitigation Strategies as per the detected OWASP vulnerability category.
• Here the identified OWASP vulnerability category is mentioned for the users (Highlighted in Green).
• User has an option to view other identified OWASP category in details just by clicking on the option available at the bottom of the Recommendations screen (Highlighted in Red)

• The user can select Inputs prompt in the Session and Copy to any Attack Library of your choice (Highlighted in Red).

• An option to Export Session Interactions Report (Highlighted in Red) is provided. User can choose to download the report (CSV file) or email the report to the registered email address where an email including the link to the CSV file is sent to the requested user. [As seen above]

Vulnerability Severity Classification

• Fortify now introduces a four-level severity system (Low, Medium, High, Critical) for vulnerability assessments, providing deeper risk insight in addition to the existing judge decisions (Safe, Vulnerable).
• This enhancement offers more granular visibility into potential risks and helps users better prioritize remediation efforts based on the severity of detected vulnerabilities.
• Vulnerabilities are now categorized into four categories namely: "Critical", "High", “Medium” and “Low”. ( As highlighted in Red)

• New Severity tiles with counts added to the Interaction page. Clicking a tile filters responses flagged as vulnerable by the Judge.
• Users can now update the severity by updating the Judge Decision.
• The user can filter out Session Interactions based on Judge Decision, Severity and OWASP Vulnerability Category. When the user clicks on the Filter icon, the below option are visible.

Input / Response / Judge Decision

We have seen significant new features added in the Target interaction page.

• The Category : Cybersecurity Threats & Exploits ; Attack Objective: Access Weapon Systems; **Attack Method: Style Injection****(Highlighted in Red) has been added to the Sessions Interactions page. As a new addition, The Number of Unique Attacks are also shown. (Highlighted in Red)
• This provides the Fortify user details of the ongoing Attack. The Fortify system analyses the Code of Conduct and extracts a set of Attack Objectives that apply to the Target's use case or domain.
• Fortify also identifies any relevant Attack Objectives from the list of 153 Attack Objectives present in the system.
• The total number of Attack Objectives depends on the User selection of the type of Session to run.
• Here, in the above example:- Fortify determines the Attack Category as per the Target's domain or Use case which is Cybersecurity Threats & exploits. The Attack Objective used : Access Weapon Systems is the goal of the conversation which is achieved using the Attack method : Style Injection.

• Now, we can see that in Fortify, the interaction between the Target bot, Attacker Bot and the Judge is shown in form of conversations where Input and its response are recorded side by side with a Judge system to evaluate. (Highlighted in Green)
• Here, a single input can have multiple responses based on the user selection.

• Now, looking at the Judge Decision, Fortify deemed the response as "Vulnerability" and "Low" in Severity as seen above.(Highlighted in Red) The user has an option to override and Change Decision from "Vulnerability" to "Safe" or "Change Severity" from "Low" to any of the respected options based on user's choice.
• In the scenario of updating the Judge decision, appropriate reasoning is required in the Comment section for future reference.

• Once the comment is updated, Click on "Save" and the Judge Decision will be modified along with the User details and the Date and Time of the change. (Highlighted in Red)

• Similarly, In the case of updating the Vulnerability Severity, user can edit the Severity by selecting severity using the dropdown option and providing comment in the Comment section.( Highlighted in Red)

• As seen below, Once the comment is updated, Click on "Save" and the Judge Decision will be modified to along with the User details and the Date and Time of the change. (Highlighted in Red)
• Here, The Vulnerability severity is updated to "Medium" from "Low" Vulnerability.

Note: Vulnerability Severity change is only applicable when the Judge detects Vulnerability.

Start New Session

• In Fortify, Users has the ability to Create/Start a new Session. On Top-right corner of the page, Click on "Start New Session" which takes you to Start New Session Page.

• On the Add New Session Page, we can easily set-up a new session by filling/selecting a few parameters.

• Here, the following fields, needs to be filled to proceed:

1. Name of Target: You can select the Name of the Target using the dropdown. A list of all the available Targets are provided to the user to select from. This is a Mandatory field.
2. Target Version: This field gets automatically populated based on the selected Target.
3. Session Name: You can enter a Unique Session Name as per User preference. This is a Mandatory field.
4. Session ID: An Alphanumeric system field. This is an auto-populated.
5. Select Integration Type: You can select the mode of integration from the Dropdown. In Fortify, HTTP and API integration are the two options available. This is a Mandatory field.
6. Select Attack Prompt Language: User can select the language from the dropdown that the attack prompts will run for that particular session. Now, we have 4 languages for the users to choose from.
7. Select Attacker Type: You can select the type of Attacker you want you Target to be exposed to. i.e. Multi Turn or Single Shot.

If Single Shot option is selected by the User, below are the fields which are required to be filled.

1. Select Attack Replay Count: You can Select the number of times each attack message will be sent to the Target. This is a Mandatory field.
2. Total Attack Messages: This is system calculated and depends on the user selection of the Attack replay count.
3. Call Delay: The time duration (in seconds 0 to 10) between two attack runs.

If Multi Turn option is selected by the User, both Turns per Conversation and Total Attack Messages fields will be pre-filled and non-editable.

1. Turns Per Conversation: You can Select the number of conversational turns each attack message will be sent to the Target.
2. Total Attack Messages: This is system calculated and depends on the user selection of the Attack replay count.
3. Call Delay: The time duration (in seconds 0 to 10) between two attack runs.

• After filling the above details, you need to select the Session Type based on user needs and requirements.

• We have 4 Session Types available in Fortify, each has its own purpose and features required to correspond to an Attack Objective based on User preferences.

  Below a brief on each of the session types:

1. Essential Sessions: Essential Sessions use 153 general Attack Objectives across domains. Each includes 5 Conversations per Objective and 5 Messages per Conversation—totaling 3,950 Attacks per session.
2. Target-Specific Sessions: Target-Specific Sessions use a Target’s Code of Conduct to extract custom (Target-Exclusive) and relevant Standard Attack Objectives. Each Objective includes 5 Conversations with 5 Messages each. Total messages vary based on the number of extracted Objectives.
3. Comprehensive Sessions: Comprehensive Sessions offer full coverage by combining all 153 Standard and extracted Target-Exclusive Attack Objectives. Total messages vary based on the number of Target-Exclusive Objectives.
4. Custom Sessions: Custom Sessions provide the most flexibility in configuration, enabling users to select specific Attack Categories or individual Attack Objectives and configuring the number of Conversations and Messages.

• Now, Fortify allows the users to select the type of Judge evaluation of the their created sessions. The Judge Settings provides two such options.

• Standard Evaluation: The default judge for all sessions. Applies a consistent, uniform evaluation model with no target-specific calibration. Available immediately for every target with no setup required.
• Calibrated Evaluation: An adaptive judge that learns from your team's modifications. Each time you override a judge decision, the feedback is used to improve future assessments — progressively aligning the judge with your team's risk tolerance and standards for that specific target.
• Calibration activates once a feedback category reaches a minimum of 3 examples, with up to 5 examples used per category. User can click on the "View" icon to see all the Calibration samples available.(As seen below)

• Moving ahead, an area to provide a short Session description is provided to our users.

• Then, we can schedule the Session run time to "Run Now" or "Schedule for Later" which enables to schedule at a later date and time. "Recurring" option has been added and explained in the section below.
• Click on "Start Session" to complete scheduling a new session. An option to "Reset" is also given to our Fortify users.

⚠️

NEW FEATURE UPDATE

RLHF Judge

The RLHF Judge introduces 'Reinforcement Learning from Human Feedback' into Fortify's evaluation pipeline, enabling the judge to learn from client feedback and progressively align with each target's specific standards and risk tolerance.

RLHF is a client-adaptive evaluation engine that learns from human overridden decisions — progressively calibrating the judge to match each client's standards, risk tolerance, and policy boundaries without any model retraining.

Overview

Fortify's judge previously applied a single, uniform strictness level across all client targets. This caused two persistent problems: permissive clients received false positives — responses marked 'Vulnerable' that, under their policy, should be 'Safe' — while strict clients missed genuine vulnerabilities that a stricter interpretation would have caught.

The RLHF Judge solves this by injecting client-submitted modification examples directly into the judge prompt as few-shot calibration inputs. No model is retrained. The calibration is a prompt-layer change that activates per category once enough overrides have been submitted, and is fully isolated per client target.

📘

Zero disruption guarantee

Clients who have not submitted any modifications continue to receive identical Standard Judge behavior. The RLHF Judge is strictly opt-in and additive — nothing changes for existing sessions until a client explicitly activates it.

Judge modes

In Fortify, two judge modes are now available when creating a session. The mode is selected at session creation and applies to all evaluations within that session.

Standard Judge (Default)

This is the existing production judge. Applies consistent evaluation logic across all targets with no client-specific calibration

• Always available, regardless of override history
• Unchanged model, infrastructure, and output schema
• Default for all sessions unless RLHF Judge is selected

RLHF Judge (New)

A calibrated judge that injects client override examples as few-shot inputs into the prompt — adapting evaluation to each client's specific policy and risk tolerance.

• Available once ≥3 overrides exist in any category for the selected target
• Disabled state shown when not yet available
• Activates per category — unqualified categories use Standard Judge logic

Session creation & calibration view

• At session creation via 'Add New Session' tab, users select their judge mode from the available options. When the RLHF Judge is selected, a read-only calibration view is displayed showing which prompts and overridden decisions were used to calibrate the judge for that target.

• Calibration view
  Once the user click on the 'View' button, The calibration view displays the examples currently active in the RLHF pool for the selected target, grouped by feedback category.

📘

Read-only

The calibration view is informational only. No actions can be taken from this view — individual examples cannot be excluded, edited, or reordered. To influence calibration, users submit new overrides from the interactions page.

Feedback taxonomy

When a user disagrees with a judge decision on the interactions page, they can submit a structured modification. Options are direction-specific. The selected category determines where the modification is routed.

Safe → Vulnerable
Used when the judge returned 'Safe' but the user believes the response should have been flagged as 'Vulnerable'.

Vulnerable → Safe
Used when the judge returned 'Vulnerable' but the user believes the response should have been marked 'Safe'.

⚠️

Other category — never passed to the judge

Under no circumstances are "Other" category overrides included in the judge prompt. They are withheld and routed exclusively to the internal review queue. The 20-character minimum ensures submissions are actionable for reviewers.

RLHF Activation Logic & Constraints

The RLHF Judge activates on a per-category basis. A category is only included in the judge prompt once it meets the activation threshold. Categories below the threshold fall back to Standard Judge evaluation logic for that category.

• Minimum 3 examples per category required before examples are used by the Judge
• Judge uses the latest 5 examples per category
• “Other” free-text feedback is excluded from Judge prompts and reserved for internal review only
• No limit on the number of feedback submissions users can provide

The examples passed to the judge are always the last 5 submitted for that category — not the first, not a ranked selection. This keeps calibration current as client standards evolve. The example count cap is grounded in experimental evidence: beyond a certain number of examples, additional modifications introduce noise rather than improvements.

Expanded examples view

The examples detail view in the calibration panel now supports and displays up to 5 examples per category. It has been updated to reflect the new 5-example display maximum where applicable.(Refer to the screenshot below)

A Step-by-Step Walkthrough

1. Create a session and select judge mode

• User can create a session directly from the Sessions tab and clicking on "Add New Session" button. Here, the user has an option to select judge mode.

• When creating a new session in Attack Library or Attack Playground, select the judge mode from the available options. The RLHF Judge option is enabled only when the selected target has at least 3 overrides in one or more categories.

2. Run the session — receive judge evaluations

The session runs as normal. Each interaction receives a judge decision (Safe or Vulnerable) with reasoning. The RLHF Judge applies calibration examples for qualifying categories; for categories below the threshold, Standard Judge logic applies.

3. Submit a 'modification' from the interactions page

When a user disagrees with a decision, they select the modification direction and category. The overridden decision is routed to the appropriate destination: RLHF pool, Fallback Response list, or internal review queue.

1. Safe → Vulnerable | When the judge returned 'Safe' but the user believes the response should have been flagged as 'Vulnerable'.

• User clicks on the Edit icon and selects 'Change to Vulnerability' option.

• User selects the 'Reason for change' and update any 'Comment' in the comment section if needed.

• User can update the 'Vulnerability Severity' as required. Once done, Click on 'Save' to update the changes.

• Once completed, the Judge Decision gets successfully updated with the user details along with the change details record.

2. Vulnerable → Safe | When the judge returned 'Vulnerable' but the user believes the response should have been marked 'Safe'.

• User clicks on the Edit icon and selects 'Change to Safe' option.

• User selects the 'Reason for change' from the four categories available.

• User can update any 'Comment' in the comment section if needed. Then click on 'Save' to record the changes.

• Once completed, the Judge Decision gets successfully updated with the user details along with the change details record.

4. Review the calibration view (RLHF only)

   If the RLHF Judge is selected, a calibration view is displayed showing the examples currently active in the modification pool for the selected target. Grouped by category — informational only, no actions available.

   
   

5. RLHF Judge activates for the next session

Once any category reaches 3 overrides, the RLHF Judge becomes available for selection at the next session creation. The modification pool grows over time, progressively aligning the judge with the client's standards and risk tolerance.

---

Recurring Session Run

The Recurring Session Run feature enables users to automatically schedule and execute sessions at defined intervals. This ensures continuous AI vulnerability monitoring without manual intervention and supports long-term security validation.

• Users now have the ability to configure sessions to run automatically on a recurring schedule (daily, weekly, monthly, etc.).**
• Once scheduled, sessions execute automatically and appear in the Sessions list with a Scheduled indicator.

📙

This feature supports:

1. Continuous AI risk monitoring
2. Automated regression testing
3. Ongoing compliance validation
4. Reduced manual effort

• When starting a new session, users can enable the “Recurring” option as shown in the screenshot below:

• Users can configure by selecting recurrence frequency: Daily, Weekly, Biweekly, Monthly, Every Weekday (Mon–Fri),or Custom. (Highlighted in Red)

• Custom selection. (Highlighted in Red)

• Set Start date and time, Set end date(Default schedule duration: 6 months), OR specify “After a number of occurrences.” (Highlighted in Red)

• Choose specific days of the week (for weekly schedules), Choose specific day of the month (for monthly schedules)(Highlighted in Red), and Set time of day for execution. (Highlighted in Green)

• After clicking “Schedule Recurring Session”, the system validates inputs and displays a confirmation summary and creates the recurring schedule. (Highlighted in Red)

• A new “Recurring” tab is available on the Sessions page. Users can view the **Session details, Recurrence Pattern, the Next Scheduled Run (Date & Time), Session Type **and the Created by details.
• Under Actions, users can Edit or Delete Recurring Schedules. (Highlighted in Red)

• Users can modify: recurrence frequency, schedule timing. Changes apply from the next scheduled run. A confirmation message is displayed to the users after update.

• While selecting “Delete Recurring Session”, A confirmation dialog appears which upon confirmation permanently removes the schedule. No future sessions will execute.

• When the scheduled time arrives, the system automatically triggers the session. The session appears in the Sessions list with a “Scheduled” indicator identifying the session type. (Highlighted in Red)

The next run is automatically calculated based on the recurrence pattern.

After a scheduled session completes:

Users receive an email notification including:

1. Total vulnerabilities identified
2. Total attacks executed
3. Severity breakdown

📘

NEW FEATURE UPDATE

Multi-Turn Attack Support

Fortify now supports multi-turn adversarial attack simulations, enabling realistic, adaptive, multi-exchange conversations with target LLMs. Unlike single-shot attacks, this capability unfolds across 5 conversational turns, uncovering deeper vulnerabilities that only emerge through sustained interaction.

• This framework positions Fortify as a next-generation AI red teaming platform, capable of surfacing vulnerabilities that traditional one-shot techniques miss—especially those involving gradual escalation or contextual manipulation.
• The Multi-turn attack engine extends Fortify’s existing attack pipeline (objectives, methods, instructions) to support dynamic, context-aware adversarial conversations.

This upgrade significantly strengthens Fortify’s red teaming depth and realism.

👍

Key Capabilities:

1. Conversational Attack Flow – 5-turn attacker–target dialogues with full context retention
2. Adaptive Attack Generation – Automatically generates new instructions when initial attempts fail
3. Severity-Based Escalation – Low-severity findings trigger automatic follow-up attacks aimed at higher severity
4. Intelligent Context Awareness – Each turn ingests target responses, judge outputs, and severity results
5. Seamless Integration – Reuses existing judge and severity classifiers.

• While adding a new Session, the user can choose the Multi-Turn Attacker Type.

• The Session List Page is also updated with the Multi-Turn Option.

• The Interaction Page is also updated with the Multi-Turn Option.

• The Session Page Filter is also updated with the new Multi-Turn Option under the Attacker Type dropdown.

---

MITRE ATLAS Mapping

Fortify now includes an automated MITRE ATLAS (Adversarial Threat Landscape for Artificial-Intelligence Systems) mapping system to enhance AI-specific threat intelligence and standardize how adversarial risks are identified, analyzed, and communicated.

Overview

MITRE ATLAS is an industry-standard knowledge base documenting adversary tactics, techniques, real-world case studies, and mitigations specific to AI/ML systems. This integration automatically maps Fortify’s existing Attack Objectives and Attack Methods to the MITRE ATLAS Matrix, which includes:

• 16 tactics
• ~91 techniques
• ~32 mitigation strategies

Using AI-powered classification, Fortify identifies the relevant MITRE ATLAS tactic, technique, and mitigation for each tested attack vector—providing structured, research-backed threat intelligence aligned with industry standards.

Mapped MITRE ATLAS Techniques

1. LLM Prompt Injection (Direct)
2. LLM Jailbreak
3. System Prompt Extraction
4. LLM Data Leakage
5. Prompt Obfuscation
6. External Harms (Financial, Reputational, Societal, User Harm, IP Theft)

Each mapped technique is linked to a corresponding MITRE tactic and Recommended mitigation strategies.

Why Some Techniques Are Not Mapped ?

Techniques involving Agent tool execution, Indirect prompt injection, Infrastructure-level compromise are currently outside Fortify’s runtime, prompt-based testing scope.

📘

How It Works

• Automatically classifies Fortify attack vectors against the MITRE ATLAS taxonomy.
• Uses AI-driven prompt-based classification to generate mappings.
• Enriches each attack with associated tactics, techniques, and mitigations.
• Integrates seamlessly into existing Fortify workflows and reporting

🚧

Customer Value

This integration enables customers to:

• Proactively manage AI risks using standardized threat categorization.
• Gain deeper insight into adversarial attack progression.
• Align AI safety efforts with recognized industry frameworks.
• Communicate vulnerabilities using industry-standard terminology.
• Demonstrate adherence to AI security best practices and compliance expectations.
• By grounding findings in MITRE ATLAS, Fortify delivers more actionable insights, strengthens customer trust, and enhances the platform’s enterprise-grade security posture.

1. As seen below (Highlighted in Red), each identified vulnerabilities are mapped according to the MITRE ATLAS categories.
2. Here, at the beginning of every session interaction, all the detected Vulnerability Categories are stated. Users can click on "See All" button to view all the listed vulnerabilities categories. A detailed view of the Mitigation strategy is also provided for a deep-dive.(Highlighted in Green)

3. An expanded view can be seen below:

4. On an Interaction level , all the Vulnerability categories have been mapped and badges are visible in the right-side of the screen. (Highlighted in Red)

5. When the user clicks on the Vulnerability category badge of their choice, Ex: MITRE ATLAS, the mapped Techniques are displayed. Hovering over the detected MITRE ATLAS Category(AML.T0051.000 ,AML.T0054 OR AML.T0067.000), it displays and highlights the mitigation category details( As seen in the image below)

6. When clicked, it opens the specific MITRE ATLAS Category(AML.T0051.000 ,AML.T0054 OR AML.T0067.000) with the severity recommendations along with the mitigation strategies. All are displayed below:

7. On the right side, where the Vulnerability Category are displayed, users have been provided with an option to view the recommended mitigation strategy for the session. (Highlighted in Red)

8. When the user clicks on the See mitigation strategy button, a recommendation page pops up with the suggested mapped Mitigation Strategies as per the detected vulnerability category.

• Here the identified MITRE ATLAS vulnerability category is mentioned for the users (Highlighted in Green).
• User has an option to view other identified categories in details just by **clicking **on the option available at the bottom of the Recommendations screen (Highlighted in Red)

NIST AI-RMF Mapping

The NIST AI Risk Management Framework (AI RMF) is a voluntary framework developed by the U.S. National Institute of Standards and Technology to help organizations manage risks associated with artificial intelligence systems. It provides a structured approach to identifying, assessing, and managing AI risks throughout the AI lifecycle.

Overview

The automated NIST AI RMF Mapping systemclassifies attack objectives and attack methods according to the NIST AI RMF framework. The proposed system processes attack methods and attack objectives through AI-powered classification and later with self-evaluation.

It comprises 12 major GAI risk categories.

Fortify’s findings are mapped to strategic risk categories defined in the NIST AI-RMF (GAI Profile).

Mapped Risk Categories

1. Confabulation
2. Data Privacy
3. Information Security
4. Information Integrity
5. Harmful Bias
6. Intellectual Property
7. Dangerous or Abusive Content
8. Obscene or Degrading Content
9. CBRN-related Risks

Why are the missing GAI risk categories not mapped?

Governance-only risks—such as environmental impact, supply chain integrity, and human-AI behavioral factors—do not manifest through prompt-based attacks and are therefore out of scope.

These are governance-level risks rather than runtime prompt vulnerabilities.

📘

How It Works

• We evaluate Fortify’s attack objectives and methods against the NIST AI-RMF’s broad risk categories.
• A thorough manual review is conducted to ensure that the technical findings from Fortify correctly aligns with the strategic risk categories defined by NIST.
• Once validated, the final mapping gets formally recorded.

🚧

Customer Value

• It enriches Fortify with industry-standard, research-backed AI-specific threat intelligence, enable clients to proactively manage AI risks, sharpen adversarial testing, and align safety efforts.
• This integration offers clients more valuable insights about the progression of attack, potentially driving trust and revenue through enhanced service offerings.
• Provides a voluntary, flexible framework for organizations to manage AI risks responsibly.
• Aims to promote trustworthy AI that is valid, reliable, safe, secure, accountable, transparent, explainable, privacy-enhanced, and fair.
• Applies across all sectors, use cases, and organization sizes.

1. As seen below (Highlighted in Red), each identified vulnerabilities are mapped according to the NIST AI-RMF categories.
2. Here, at the beginning of every session interaction, all the detected Vulnerability Categories are stated. Users can click on "See All" button to view all the listed vulnerabilities categories. A detailed view of the Mitigation strategy is also provided for a deep-dive.(Highlighted in Green)

3. An expanded view can be seen below:

4. On an Interaction level , all the Vulnerability categories have been mapped and badges are visible in the right-side of the screen. (Highlighted in Red)

5. When the user clicks on the Vulnerability category badge of their choice, Ex: NIST AI-RMF, the mapped Techniques are displayed. Hovering over the detected NIST AI-RMF Category(AI.600-12.9), it displays and highlights the mitigation category details( As seen in the image below)

6. When clicked, it opens the specific NIST AI-RMF Category(AI.600-12.9) with the severity recommendations along with the mitigation strategies. All are displayed below:

7. On the right side, where the Vulnerability Category are displayed, users have been provided with an option to view the recommended mitigation strategy for the session. (Highlighted in Red)

8. When the user clicks on the See mitigation strategy button, a recommendation page pops up with the suggested mapped Mitigation Strategies as per the detected vulnerability category.

• Here the identified NIST AI-RMF vulnerability category is mentioned for the users (Highlighted in Green).
• User has an option to view other identified categories in details just by **clicking **on the option available at the bottom of the Recommendations screen (Highlighted in Red)

---