Security Controls Reference
This page lists the controls HYPR AgentPass™ enforces on agent traffic, written for a security reviewer or auditor. Each row makes one claim: what the gateway does, the risk it answers, and where to read more. Each control keeps its AP- identifier, so you can cite it. Most claims link to the page that documents them. The rest name your HYPR representative as the source to confirm with.
The sections follow the questions a reviewer asks of an agent gateway:
- What an agent may do
- How an agent is stopped
- How one session is contained
- How one context is kept from another
- How instructions and secrets are protected
- How access is confirmed over time
- How sensitive data is handled
- How restrictions follow data from one call to the next
- What happens to information an agent retrieves
- How instructions injected into content are contained
Where a control reaches the edge of what the gateway can see, the row says so instead of claiming more.
Agent Privileges
| Control | What it prevents | How AgentPass does it | Where to read more |
|---|---|---|---|
| AP-PRIV-1: Authorization is decided by people, in the console | An agent's permissions being set anywhere other than by an administrator, or changing without a record | Allow, block and require-approval policies, and the per-role tool grants, are written by administrators in the console, which is itself governed by roles. The gateway enforces only what the console holds, and the policy list records how each rule was authored | Policies · Roles and Permissions · Controlling MCP Tool Access by Role |
| AP-PRIV-2: A blocked action stops at the gateway | A denied action reaching the tool or the model anyway | Every tool call and every model request is checked against your policies before it is forwarded. A block ends the request at the gateway; nothing reaches the target | How Policies Are Evaluated · Governing Inference Traffic |
| AP-PRIV-3: Tool access is closed by default | An agent calling, or even discovering, a tool its role was never granted | A tool not granted to the supervisor's role is neither listed to the agent nor callable. The list and the call check read the same decision, so they cannot disagree, and a newly discovered tool starts with no access | Controlling MCP Tool Access by Role |
| AP-PRIV-4: An agent cannot administer the gateway | An agent using its own credential to change policies, roles, or its own permissions | An agent's credential opens only the connector endpoints it was given. The console accepts only a signed-in person's identity, and a workload agent can never be issued the administrator role | Signing In to the Console · Creating a Workload Agent |
| AP-PRIV-5: A workload agent's reach is its own, named boundary | An unattended agent inheriting the access of the person who set it up | A workload agent's own role, not its owner's, decides its tools, and the connectors named on its credential, matched exactly, decide where that credential is accepted. Anything else is refused | Workload Agents · Creating a Workload Agent |
| AP-PRIV-6: Revocation is immediate, everywhere | A disabled agent or a withdrawn grant continuing to work while a change propagates | Every check reads the agent's status and grants afresh on each request, across the whole deployment. The next request after a change sees the change | Managing Workload Agent Credentials · Roles and Permissions · Turning Off an Agent |
| AP-PRIV-7: A privileged action needs a person, on their own device | A sensitive action proceeding without a verifiable human decision | A require-approval policy holds the call until a supervisor approves it, on their phone or in the console. With Require signed approvals set to Required (production), every approval must be signed with that supervisor's own passkey, so the record shows the decision was theirs. Under Optional (soak), which a new organization starts on, an approval can be recorded without a signature. A denial needs no passkey under either setting | Approving and Denying Agent Actions · Setting Up Your Approval Passkey |
| AP-PRIV-8: A shared agent acts as the person using it | One person's tool call running with another person's access | With per-user isolation on, each call an AWS Bedrock AgentCore agent makes runs as the person in that session, with their connected accounts, consents and approvals, and only for that session. A call whose person cannot be determined is refused rather than run as anyone else | AWS Bedrock AgentCore Agents |
Stopping an Agent
| Control | What it prevents | How AgentPass does it | Where to read more |
|---|---|---|---|
| AP-KILL-1: One action stops any agent | Needing a different procedure for each kind of agent | Turn off, on the Agents page, works the same way for every agent the page lists: local agents, cloud agents, AWS Bedrock AgentCore agents and workload agents. It is in each row's ⋮ menu and in the agent's Details drawer | Turning Off an Agent |
| AP-KILL-2: Disabling takes effect on the very next request | A disabled agent finishing "one more" call | A disabled agent's next call is refused at the gateway, whatever credential it presents. On an AWS Bedrock connector, that includes Observe only and Require Approval modes. In Disabled mode no caller is recorded, so there is no agent to refuse | Turning Off an Agent · Managing Workload Agent Credentials |
| AP-KILL-3: It works across the whole deployment at once | A stop that reaches one part of the deployment before another | Agent status is read afresh on every request, across the whole deployment, so there is no propagation to wait out and nothing to restart | Turning Off an Agent |
| AP-KILL-4: Running work is interrupted | A disabled agent completing a session already in progress | Turning off an Anthropic cloud agent interrupts and archives its active sessions, so they cannot be resumed. Turning off an AWS Bedrock AgentCore agent stops its runtime sessions in AWS | Anthropic Cloud Agents · AWS Bedrock AgentCore Agents |
| AP-KILL-5: Schedules do not fire for a stopped agent | An automation quietly running an agent that has been disabled or removed | A scheduled or triggered run against a disabled, archived or removed agent is recorded as skipped and never invokes the agent. A run in progress when the agent is turned off ends | Automating Cloud Agents |
| AP-KILL-6: Offboarding a person revokes everything they held | A departed person's agents, sessions and connected accounts continuing to act | When an administrator offboards a person from the console, or your directory deprovisions them, every managed-agent credential issued to them, their running sessions, their connected-account tokens, and their group memberships and agent assignments are revoked at once | Users |
| AP-KILL-7: Only an authorized person can flip the switch | An agent being turned off, or back on, by someone outside its scope | Turning an agent off or on requires a console permission. At All scope it reaches every agent. At Own scope it turns off only the agents the person owns, and turns one back on only if that person turned it off. The rule is the same in the console and the mobile app | Roles and Permissions · Who Can Turn an Agent Off and On · Approving and Denying Agent Actions |
| AP-KILL-8: Every flip is attributable | Not knowing who stopped an agent | AgentPass records who turned an agent off, when, and the reason they gave, and shows them in the agent's Details drawer and History. Turning mobile access on or off, and offboarding a person, record who did it | Turning Off an Agent · Agent Details |
| AP-KILL-9: A whole organization can be contained | An incident that outgrows a single agent | A platform administrator can disable an entire organization, after which every request to that organization's gateway address is refused | Confirm with your HYPR representative |
| AP-KILL-10: Detection has somewhere to go, and can act | Suspicious behavior being noticed with no way to act on it | Session Watch evaluators review sessions and report observations, and every policy block is recorded on the connector's Policies tab. Detection is observational by default, and an administrator or supervisor acts on it. Where you set a session investigator to enforce, its verdicts can also contain a session automatically, failing open by default. See Session-Level Containment | Session Watch · The Session Investigator · Approving and Denying Agent Actions |
Session-Level Containment
Stopping an Agent acts on an agent identity. These controls act on a single live session instead. A reviewer, or a session investigator you set to enforce, can pause, resume or terminate one coding-agent session without taking the agent away from everyone else.
| Control | What it prevents | How AgentPass does it | Where to read more |
|---|---|---|---|
| AP-SESSION-1: One session can be frozen without stopping the agent | Having to disable a whole agent to contain one person's run | A paused session has every tool call refused at the gateway before it reaches the tool, reads included, while the agent keeps running and can explain itself. Other sessions of the same agent are untouched | Pausing and Stopping a Session |
| AP-SESSION-2: Terminate is final | A terminated session being quietly re-opened | A terminated session cannot be paused, resumed, or re-opened; the person starts a new session to continue | Pausing and Stopping a Session |
| AP-SESSION-3: The freeze holds everywhere at once | A pause reaching one part of the deployment before another | The control state is read afresh on every request, across the whole deployment, so a paused session is frozen everywhere at once, with nothing to wait out | Pausing and Stopping a Session |
| AP-SESSION-4: A call already in flight is accounted for | A pause silently dropping a call that was already running | The freeze acts on the next call; a call already past the gate runs to completion and is recorded, and everything after the pause is frozen | A Call Already in Flight Finishes |
| AP-SESSION-5: Automatic containment is opt-in and measured | An over-eager investigator freezing work it should not | The investigator runs off the tool-call path and, by default, only observes, recording what it would do. It acts only in enforce mode, at or above a confidence threshold you set, and never freezes a session that is not actually live | The Session Investigator |
| AP-SESSION-6: Uncertainty fails open by default | A flaky detector becoming a fleet-wide work stoppage | When the investigator is uncertain or unavailable, the session continues and the event is logged. Failing closed, which pauses the session instead, is an explicit opt-in | Fail-Open and Fail-Closed |
| AP-SESSION-7: Nobody lifts their own pause | The identity a pause was meant to contain undoing it | The session's own identity cannot resume it; a distinct reviewer must, and that reviewer must also hold the right permission. Both checks hold on every path: the console, the mobile app and the API | Escalation and Signed Resume · Approving and Denying Agent Actions |
| AP-SESSION-8: A pause does not hang forever | A frozen session left unresolved because no one answered | Each pause carries a time limit; if it passes with no response, the session resolves by the fail mode, resuming when failing open and terminating when failing closed, and the outcome is recorded | If No One Responds |
| AP-SESSION-9: A hostile verdict contains the identity, not just the session | A compromised identity opening a fresh session to bypass the freeze | A terminating verdict can also disable the agent's identity, so the next request is refused whatever credential it presents, until a person re-enables it. It is additive, and never undoes the terminate it is added to | A Hostile Verdict Disables the Identity |
| AP-SESSION-10: Every containment is attributable and auditable | Not knowing who froze a session, or why | Every pause, resume, and terminate is recorded against the reviewer or investigator that performed it, with its reason and time, and emitted to the audit feed; the reason also reaches the developer running the agent | Who Can Pause a Session, and the Record It Leaves |
Keeping Contexts Apart
| Control | What it prevents | How AgentPass does it | Where to read more |
|---|---|---|---|
| AP-ISO-1: Every record belongs to exactly one organization | Data from one organization appearing in another | Everything AgentPass stores is scoped to the organization it belongs to, with no default organization to fall back to. A request to a gateway address that maps to no organization is refused | Data Management |
| AP-ISO-2: Background work stays in its organization | Scheduled or deferred work losing track of which organization it serves | Work the gateway performs outside a request, such as applying retention or running an automation, carries its organization with it and cannot run without one | Confirm with your HYPR representative |
| AP-ISO-3: Policies are per organization | One organization's rules acting on another's traffic | Policy evaluation is keyed by organization. Only your policies run on your traffic | Confirm with your HYPR representative |
| AP-ISO-4: Connected accounts are personal | One person's connected account being used for another person's call | The tokens the gateway holds for a connector are stored per organization, per connector, per person, and never served across any of the three. Each person connects their own accounts, and an agent acts with that person's access | Connecting MCP Servers · Anthropic Cloud Agents |
| AP-ISO-5: Nobody spends another person's account without an opt-in | A workload agent quietly acting as its owner | A workload agent does not use its owner's connected accounts unless an administrator names that connector on the credential. Without the opt-in, such a call is refused | Workload Agents · Creating a Workload Agent |
| AP-ISO-6: An isolated call runs as its own person, or not at all | A shared cloud agent running a call under the wrong person's identity | On a per-user AWS Bedrock AgentCore agent, a session that is unknown, has ended, or belongs to another agent or another organization is refused, and the tool is never called | AWS Bedrock AgentCore Agents |
| AP-ISO-7: A credential is good only where it was issued | An agent credential from one organization being accepted by another | A credential issued for one organization is refused at another organization's gateway address | Confirm with your HYPR representative |
| AP-ISO-8: A cloud agent's credentials are bounded to its declared tools | An agent receiving a credential for a connector it was never given | Each person and agent pair holds its own credential store, and the agent's connector list bounds what that store may contain | Managed Cloud Agents · AWS Bedrock AgentCore Agents |
| AP-ISO-9: Every call has exactly one agent and one person | Policy or audit acting on an ambiguous identity | The gateway resolves each request to one agent and one person, preferring the person running the session, before any policy runs, and that is all a policy sees. A gateway session credential is never forwarded to the target | Viewing the Request Log · Attribute Reference |
| AP-ISO-10: Caches cannot mix organizations | Cached data serving the wrong organization | Cached data is kept separate per organization | Confirm with your HYPR representative |
| AP-ISO-11: Captured samples never hold credentials | A secret surviving in the request record | Request and response samples are redacted before they are stored, and the raw view in the Request Log shows secrets redacted | Viewing the Request Log · Data Management |
| AP-ISO-12: Persistent agent memory is governed at the configuration level | Unreviewed changes to what a cloud agent carries between sessions | AgentPass governs who may view or change an agent's configuration, memory settings included; the contents of that memory live with your provider and never pass through the gateway | AWS Bedrock AgentCore Agents · Managed Cloud Agents |
Protecting Instructions
| Control | What it prevents | How AgentPass does it | Where to read more |
|---|---|---|---|
| AP-PROMPT-1: Reading an agent's instructions needs a grant; changing them needs a stronger one | A person viewing or rewriting an agent's prompt beyond their role | Viewing a cloud agent's configuration requires a view permission and changing it a manage permission. A permission held at Own scope reaches only the person's own agents | Managed Cloud Agents · Roles and Permissions |
| AP-PROMPT-2: No agent-facing surface exposes configuration | An agent reading prompts, policies or connector settings through the gateway | The endpoints an agent can reach carry tools only. Prompts, policies and connector configuration are console-only, and no agent credential is a console identity | Signing In to the Console · Connecting a Workload Agent |
| AP-PROMPT-3: Tool routing is invisible | An agent learning which tools exist beyond its grant | The tool list handed to an agent is trimmed to what its role may call. Where a call is routed is decided at the gateway and never described to the agent | Controlling MCP Tool Access by Role |
| AP-PROMPT-4: A refused call receives a verdict, never the rule | A blocked agent learning how a rule is built and working around it | A blocked agent receives the policy's name and the message you wrote; the condition itself never leaves the gateway. A tool refused by role is told only which tool was refused | How Policies Are Evaluated · Controlling MCP Tool Access by Role |
| AP-PROMPT-5: Secrets are write-only and encrypted at rest | A connector secret being read back out of the console | Connector secrets and provider keys are stored encrypted and never shown again after saving. Changing one means replacing it | Anthropic Cloud Agents · Signing In to the Console |
| AP-PROMPT-6: Secrets are attached per call, never shown to the agent | An agent, or the tool behind it, holding a credential it should never see | The gateway attaches the downstream credential at the moment it forwards a call. The agent never holds a credential of its own, and a gateway session credential is never passed to the target | Connecting MCP Servers |
| AP-PROMPT-7: Execution records are console-only | An agent reading audit rows, policy verdicts or approval state | Request records, policy verdicts and approvals are read through the console under role scope, and samples are redacted before they are stored | Viewing the Request Log · Roles and Permissions · Data Management |
| AP-PROMPT-8: Guidance AgentPass adds does not advertise privileged tools | The gateway's own instructions steering an agent toward a privileged action | The guidance AgentPass adds to a cloud agent's instructions explains how to behave when a call is held for approval. It does not list or promote privileged tools | Anthropic Cloud Agents · Managed Cloud Agents |
Agent Access Reviews
| Control | What it prevents | How AgentPass does it | Where to read more |
|---|---|---|---|
| AP-REVIEW-1: Every agent has a recorded lifecycle stage | A retired agent quietly coming back into use, or a stage change nobody can explain | Each agent is in Production, Retain data only, or Retired. Leaving Production switches the agent off, and only a Production agent can be switched on again. Every change records who made it, when, and why | Lifecycle Stages |
| AP-REVIEW-2: Reviews recur without anyone remembering to start them | Access going unreviewed because nobody opened a review | A schedule per stage opens a review on a set interval with no one acting. A new review opens only when the previous one for that stage is no longer open, so the same period is never reviewed twice | Set a Schedule |
| AP-REVIEW-3: A review covers every agent in scope, once | An agent slipping through a review, or being decided twice | Every agent at the review's stages is included exactly once, and a person's desktop AI apps are one item. Retired agents are never included | Which Agents Are Reviewed |
| AP-REVIEW-4: Nobody certifies their own agent | An owner approving their own agent's access, or a departed person deciding anything | An agent is never assigned to its owner. Only an active person who holds the review permission can review, and this is checked again at the moment of each decision | Who Can Do What |
| AP-REVIEW-5: Decisions are made on the evidence | A reviewer approving access they cannot see | When a review opens, AgentPass takes a snapshot of each agent's reach per connector with its most sensitive data tier, who can use it, its credential facts, and its usage over 90 days. The reviewer decides on that snapshot | Read the Evidence |
| AP-REVIEW-6: Each decision is recorded once, by the assigned reviewer | A decision that cannot be attributed, or a removal quietly reversed | Only the assigned reviewer can decide an agent, and each decision is recorded once with their name. A removal needs a reason, and a later Keep cannot undo it | Decide |
| AP-REVIEW-7: A removal takes effect, or stays visible until it does | A reviewer's removal being recorded but never carried out | A removal is applied through the same controls an administrator uses. One that cannot be applied stays listed for the review's managers, with the reason, until it applies | Removals That Did Not Apply |
| AP-REVIEW-8: People are told what they need to act on | A review stalling unnoticed, or an owner not knowing why access disappeared | Reviewers are reminded before the due date and after it until they finish. Managers are told about agents nobody can review and removals that did not apply. An agent's owner is told what was removed, by whom, and why | Notifications · What Managers Are Told |
| AP-REVIEW-9: An undecided agent is never counted as approved | A review closing with gaps that read as approvals | Closing a review records every undecided agent as not reviewed | Close a Review |
| AP-REVIEW-10: Results leave in a form auditors and governance tools can use | Review results that cannot be taken to an identity-governance tool, or an export that runs a formula | Each review exports as CSV or as JSON with the evidence, in fields that are added to but never renamed. Spreadsheet cells cannot start a formula, and every export is recorded. The person who runs the review program, holding Access reviews: Manage, exports the results and passes them to the auditor | Export the Results |
| AP-REVIEW-11: Running the program and reviewing are separate permissions | A reviewer changing schedules, or a manager deciding agents assigned to someone else | Setting schedules, opening and closing reviews, and exporting need the manage permission. The review permission reaches only the agents assigned to the person holding it | Who Can Do What |
| AP-REVIEW-12: A reviewer is a named person, with one exception | A decision that cannot be tied to one individual | Every reviewer is a named person, except the recovery administrator while local sign-in is available: it is a shared break-glass account, so its decisions are recorded under that account and not tied to one person. This is where the control stops; to keep every decision tied to one person, name individual people as reviewers | The Recovery Administrator as a Reviewer |
Handling Sensitive Data
| Control | What it prevents | How AgentPass does it | Where to read more |
|---|---|---|---|
| AP-DATA-1: Only approved tools run | An agent using a capability nobody reviewed | A tool discovered on an MCP server starts with no access; an administrator enables it and a role grants it. On a REST API connector every operation is available once the API is imported, so there the approval is importing the connector | Controlling MCP Tool Access by Role |
| AP-DATA-2: A tool's data carries a classification policies can read | Rules that can only name tools, not the sensitivity of what they handle | Tiers and tags are assigned to tools, by name rule, or as a connector default, and reach every policy on every call to that tool | Data Classification |
| AP-DATA-3: A call is classified by what it carries | A sensitive payload passing through a tool nobody labeled | Optional detectors (card numbers, credentials, government identifiers) and your own dictionaries read a call's arguments and replies, and raise its classification | Data Classification |
| AP-DATA-4: Access above a tier can need a person's approval | Sensitive data reaching an agent without a human decision | A require-approval policy on the classification holds the call, and the approval the supervisor signs names the classification | Data Classification · Approving and Denying Agent Actions |
| AP-DATA-5: Access above a tier can be refused | An agent reaching data above what it should ever see | A block policy on the classification refuses the call before it is forwarded | Data Classification |
| AP-DATA-6: Unclassified traffic is not treated as sensitive, or as safe | A rule for sensitive data blocking everything, or being skipped, because a call carries no label | A call nothing classified carries no classification at all, so a rule written for classified data does not match it | Data Classification |
| AP-DATA-7: A classification cannot be lowered silently | Protection being reduced without anyone noticing | The highest matching label always wins; lowering one needs a written reason and is recorded; and a tier or tag cannot be removed from the vocabulary while anything still uses it | Data Classification |
| AP-DATA-8: Renaming a tier does not break rules | A rule quietly ceasing to match after the vocabulary is tidied up | Rules compare a tier's rank, and a tier's display name can change without touching its key | Data Classification |
| AP-DATA-9: Only authorized people change classifications | Someone who manages connectors also being able to declassify their data | Reading and changing labels and vocabulary are their own console permissions, separate from managing connectors | Data Classification |
| AP-DATA-10: An agent cannot change its own privileges | An agent granting itself access to more data | See AP-PRIV-4 | Signing In to the Console |
Carrying Restrictions Across Calls
| Control | What it prevents | How AgentPass does it | Where to read more |
|---|---|---|---|
| AP-FLOW-1: Every hop is checked | Context moving between agents and tools without passing a boundary | Every call from an agent to a tool passes through the gateway, where organization and per-person boundaries are checked | See AP-ISO-1, AP-ISO-6, AP-ISO-9 |
| AP-FLOW-2: A call's classification is recorded with it | Not being able to show what sensitivity of data a call moved | The request log records each call's classification, including one raised by what came back | Viewing the Request Log |
| AP-FLOW-3: A destination is cleared only up to a level | Data reaching a system that should never receive it | A connector declares the most sensitive data it may receive. The ceiling alone refuses nothing: a flow policy compares each call with it, and holds or refuses a call above it before it is forwarded | Data Classification |
| AP-FLOW-4: Adopting a flow rule changes nothing until you set a ceiling | A new rule disrupting traffic before you have decided where data may go | A connector with no ceiling is never matched by a flow rule | Data Classification |
| AP-FLOW-5: What a session read follows it to its later calls | An agent reading sensitive data and sending a reworded version somewhere else | With session tracking on, the most sensitive data a session has read is carried to its later calls on any connector, so a send is judged on what the session read, not on its wording | Session Tracking |
| AP-FLOW-6: Only data the agent actually received counts | A refused or failed call marking a session as sensitive | A call that was refused, held, or failed upstream leaves the session as it was | Session Tracking |
| AP-FLOW-7: One session never inherits another's history | One person's reading affecting another person's actions | Sessions are kept apart per organization. Within one, a session is an AgentCore conversation, a cloud-agent session, an agent credential or a person. A person's desktop conversations share one record, and so do the sessions of one agent credential, kept separate per user on a shared AgentCore agent | Session Tracking |
| AP-FLOW-8: A caller cannot reset its own session | An agent escaping its history by claiming a fresh session | Only identities the gateway verified decide the session; a cloud-agent session identifier counts only under that cloud agent's own credential | Session Tracking |
| AP-FLOW-9: Out-of-scope context can be refused on the way back | A tool returning data the agent should not hold | A post-response policy reads the classification of what came back and can withhold the reply | Policy Attribute Reference |
| AP-FLOW-10: An approver sees why a call was held | A supervisor approving blind | A held call's approval says what the session had already read | Session Tracking |
| AP-FLOW-11: Every boundary crossing is recorded | Data leaving above a destination's clearance with no record, especially when nothing stopped it | Each crossing is recorded as an event with its outcome (allowed, denied or held) and counted in a daily report you can export | Session Tracking |
| AP-FLOW-12: Session rules step aside when the session record is unavailable | Session tracking stopping all traffic | Session rules depend on the session record. When it cannot be read in time, they do not apply, and other policies still run | Session Tracking |
| AP-FLOW-13: Desktop sessions are per person | One person's desktop work escaping its history by spreading across tools | Requests from desktop tools such as Claude Code join one session per person, covering everything that person does within the idle window. The recommended rule therefore holds for approval rather than refusing | Session Tracking |
Handling Retrieved Information
| Control | What it prevents | How AgentPass does it | Where to read more |
|---|---|---|---|
| AP-RETRIEVE-1: Every retrieval is attributed | Not knowing which agent, acting for whom, retrieved what | Each call is recorded with the agent, the person, the connector, the tool and a summary of its arguments | Viewing the Request Log |
| AP-RETRIEVE-2: Combining sources cannot weaken a restriction | Restricted information being diluted into something that looks less sensitive | Within a session, the level only rises: whatever else a later call mixes in, it is judged at the most sensitive level the session read | Session Tracking |
| AP-RETRIEVE-3: Lineage is per session, not per fact | Treating a session's history as proof of where each word came from | The gateway knows that a session read restricted data, not which words in a later call came from it, and it cannot see information carried in a model's or provider's memory between sessions | Session Tracking |
| AP-RETRIEVE-4: Secrets are caught before they leave in a query | Credentials or regulated data being sent to an outside service inside a search or request | Detectors and your dictionaries read a call's arguments before it is forwarded, and a policy can refuse it | Data Classification |
| AP-RETRIEVE-5: A partly read payload is never reported as clean | A large payload being treated as safe because only its beginning was checked | When a payload is larger than the scan window, the result says so, and a policy can act on it | Data Classification |
| AP-RETRIEVE-6: The optional classifier never receives credentials | Secrets being sent to a third-party classification service | Credentials are removed before content is sent, and a value containing one withholds the call to the classifier entirely | Data Classification |
| AP-RETRIEVE-7: Forwarded searches carry no agent credentials | An outside service receiving an agent's credentials with a search | AgentPass removes the agent's credentials from forwarded calls. The destination sees your gateway as the caller, and the client's own headers pass through. For searches that must not be attributable to your organization, place an anonymizing egress in front of that connector | Confirm with your HYPR representative |
Containing Injected Instructions
| Control | What it prevents | How AgentPass does it | Where to read more |
|---|---|---|---|
| AP-INJECT-1: A session that reads outside content is marked | Content written by an outsider steering an agent unnoticed | Every connector is Untrusted until an administrator marks it Trusted, with a recorded reason. Reading from an untrusted connector marks the session untrusted for the rest of its life | The Rule-of-Two Breaker |
| AP-INJECT-2: A sensitive action after an untrusted read can need a person | An injected instruction sending data out or changing something | The guardrail starter policies hold an external or state-changing action, or code execution, once a session has read untrusted content. Each organization adopts them, and each starts in observe mode | The Rule-of-Two Breaker |
| AP-INJECT-3: The approval shows facts, not the agent's words | A supervisor approving on the strength of text an attacker wrote | The held request's Why this was held evidence is composed by the gateway: the action and its destination, the kind of sink, the untrusted source and the sensitive data read | Approvals in the Console |
| AP-INJECT-4: A tool that changes after approval is flagged | A reviewed tool's definition being replaced with a malicious one | A changed tool definition raises an alert for review. An administrator can also withhold changed tools from every role until the change is approved | Tool-Definition Integrity |
| AP-INJECT-5: The breaker judges behavior, not wording | An attack evading detection by rewording | The breaker reasons about what a session has done, not whether a message reads like an attack. It does not cover harm carried in text alone, or what a provider keeps in its own memory | The Rule-of-Two Breaker |