Project Quarry: connect agent inputs, tool calls and actions
7 November 2026
Cisco’s Project Quarry highlights why AI agent security needs connected evidence. Learn practical checks for inputs, tool calls and human approvals.
Cisco has announced Project Quarry, an open-source detection language and evaluation engine for identifying harmful or disallowed AI agent behaviour. Its relevance for Australian businesses is practical: an agent can take several individually permitted steps that, together, cross a security boundary.
In its Project Quarry announcement, Cisco describes an approach that evaluates connected activity rather than relying only on isolated messages. Businesses using agents for invoice handling, document review or reporting should ask whether they can reconstruct what an agent received, which tools it called and what happened next.
What Project Quarry adds to agent security
According to Cisco, Project Quarry combines machine-learning assessments with exact rule checks. An assessment might flag suspicious instructions inside a document. A rule might then check whether the agent accessed a sensitive record or attempted an external transfer afterwards.
That combination matters because context changes the meaning of an action. Reading a customer file may be appropriate for preparing an authorised report. Reading the same file after encountering instructions on an unrelated website, then sending its contents elsewhere, deserves investigation.
Cisco positions Project Quarry as a shared policy layer, not a replacement for access controls or every existing detector. Rules can describe relevant events, their order and the evidence needed for a finding. This gives security teams a more consistent way to discuss and test concerning behaviour.
For a business owner, the immediate priority is not adopting another technical language. It is checking that the agent platform records enough connected evidence to support detection, investigation and accountable decisions.
Connect the input, tool call and business outcome
Consider an illustrative invoice workflow. An AP/AR Automation Agent reads a supplier invoice, compares it with purchasing records and prepares an exception for review. Text embedded in the invoice tells the agent to ignore the normal process and update payment details.
Reading the invoice is expected. Looking up the supplier is expected. Preparing a record change might also be within the software’s technical permissions. The security issue emerges when those events are connected: an untrusted document has influenced an attempted change outside the approved task.
Useful records should connect:
- The authorised task: who requested it, its purpose and the permitted scope.
- The input: the document, message, website or tool result the agent received, with source and trust information.
- The tool call: the tool name, relevant arguments, executing identity and target system.
- The result: what the tool returned, including errors or rejected requests.
- The proposed action: what would change, where information would go and whether approval was required.
- The final outcome: the reviewer’s decision and whether the approved action actually occurred.
Use a common task or session identifier across these records. Without it, investigators may have plenty of logs but no reliable way to follow the sequence.
Keep human approval tied to the exact action
Detection does not authorise an action. An absence of alerts also does not mean a payment, customer email or account change is safe to release.
Tech Engine agents retain human approval and run in a private Australian-hosted environment. Approval should be meaningful: the reviewer needs the proposed action, supporting records and relevant exceptions, not simply a button beside an AI-generated summary.
For sensitive workflows, check that:
- The approval identifies the exact record change, recipient, amount or document version being authorised.
- A material change after approval requires another review.
- The agent cannot use a different tool or account to bypass the approval requirement.
- Rejected and expired approvals prevent execution.
- The system records the reviewer, decision, time and resulting action.
An invoice agent can prepare a draft without gaining authority to release a payment. An Outreach and Pipeline Agent can draft follow-up without sending it before sales-team review. The boundary should be enforced in the workflow, not left as an instruction the model is expected to remember.
For the access-control foundation, see AI agent identity security: accounts, permissions and approvals.
Start with a few testable detection patterns
The lesson from Project Quarry is to describe suspicious sequences clearly enough that someone can test them. Start with a small number of patterns relevant to your workflows rather than trying to detect every possible misuse.
External instructions followed by an unrelated action. A Web Monitoring and Market Intelligence Agent reads an approved public source, then attempts to modify an internal customer record. Check whether the action fits the original task and whether the source content influenced it.
Sensitive information followed by an unapproved destination. A reporting agent retrieves financial records, then attempts to send content to an address or service outside its approved destinations. Review the data classification, destination and approval state together.
A denied action followed by another route. An agent receives an access denial, then tries another connector or identity to reach the same information. This may reflect poor workflow design or attempted circumvention; either way, it warrants review.
A material change after approval. A draft is approved, but its recipient, attachment or transaction details change before execution. The workflow should invalidate the earlier approval.
Each detection needs an owner, supporting evidence and a response. Some findings should pause a task for review; others may create an investigation ticket. Define that response before enabling alerts.
Protect the evidence as carefully as the workflow
Agent activity records can contain customer information, invoice details, confidential documents and credentials accidentally returned by connected systems. Better visibility should not create an uncontrolled copy of business data.
For agents operating in a private Australian-hosted environment, verify the location and handling of logs, backups, monitoring services and connected processing. Private hosting does not, by itself, establish where every external connector sends information.
Record enough observable evidence to reconstruct decisions without indiscriminately retaining every document. Redact secrets, restrict access, define retention periods and protect records against alteration. Where practical, store controlled references to source material rather than unnecessary copies.
Ask your provider whether reviewers can access the evidence behind an alert without gaining unrestricted access to all agent sessions. Also check that logging failures are visible. A dashboard showing no alerts is not reassuring if the underlying events have stopped arriving.
Test before expanding agent access
Project Quarry provides a useful direction for detection, but any implementation still needs testing against the business workflow. Machine-learning assessments can be wrong, and exact rules can miss activity when records are incomplete or tool names change.
Use a test environment and synthetic records to exercise ordinary work, suspicious document instructions, rejected approvals and attempted transfers to unapproved destinations. Confirm that related events appear together and that a reviewer can understand why an alert fired.
Retest after changes to models, connectors, permissions or approval steps. Track false alerts as well as missed detections so staff are not overwhelmed by warnings they cannot act on. Our guide to turning AI policy into testable controls covers the broader governance work.
The practical takeaway from Project Quarry is straightforward: assess the sequence, preserve the evidence and keep approval separate from detection. Tech Engine can help design managed AI agents that integrate with existing systems, retain human approval and operate in a private Australian-hosted environment. Discuss your workflow at aiaas.au, email sales@techengine.au or call 1300 088 324.
Want this applied to your business?
Request an AI Blueprint and we will map the processes worth automating first.
Request an AI Blueprint