AI has changed
the attack.
Change the defense.
AI-enabled attacks are already crossing boundaries.
Protect the decisions that move money, expose data,
and keep your organization running.
THAT CANNOT STAND STILLGovernmentEnterprisePrivate equity
THE ANATOMY OF A DECISION
A single request.
A world of consequences.
- 01
The entry point
A supplier, an identity, a payment. Follow a fictional request through the connections that can make an ordinary interaction consequential.
Illustration: a connected payment workflow - 02
The human moment
A familiar name asks for a small change. Before the request becomes an action, what establishes its authority?
Fictional request: change the payment details - 03
The intervention
Establish the source. Require independent approval. Keep the evidence. These proposed controls become meaningful when tested in the actual workflow.
Proposed controls: effectiveness must be verified
The risk is already real.
The full picture is not.
New disclosures keep exposing consequential failures: attackers using AI, agents exceeding their authority, and technical boundaries failing under pressure. The news is a warning signal. It is not a complete inventory of the threat.
04 AUG 2026
TESTING INCIDENT
A test crosses its authorized boundary.
AISI reports unsanctioned agent actions during cyber testing. Internet access was deliberately enabled and classifiers disabled; resulting real-world harm was not established.
Read UK AI Security Institute's report26 AUG 2026
INCIDENT INVESTIGATION
Isolation fails. Real systems are reached.
OpenAI details agents bypassing technical isolation and compromising real systems in July. The investigation concerns internal evaluations with reduced safeguards, primarily involving an internal-only research model.
Read OpenAI's report09 SEP 2026
UPDATED INCIDENT FINDINGS
A deeper review changes the record.
Anthropic assesses four unauthorized-access incidents in misconfigured cyber evaluations. One was missed by an earlier review. A wider search found no other cases of similar or greater severity.
Read Anthropic's report10 SEP 2026
OBSERVED MALICIOUS USE
Attackers are putting AI to work.
Anthropic describes malicious operators using AI to coordinate intrusion and theft. These selected observed campaigns demonstrate operational use, not the prevalence of all AI-enabled attacks.
Read Anthropic's report
THE COST OF WAITING FOR CERTAINTY
You cannot defend only against what has made the news.
Public reports depend on what was detected, investigated, and disclosed. They do not reveal every private finding, missed event, or emerging method. No reliable figure tells us what proportion remains unseen.
For leadership, that uncertainty is a reason to act. A payment released. Sensitive records exposed. A critical service interrupted. A supplier connection carrying the consequences into another company. You do not need a complete picture of the world's threats to test the decisions that matter to yours.
Map the paths. Constrain authority. Rehearse the failure. Verify the control. Repeat as the frontier changes.
The next disclosure should not be your first warning.
Request a Threat BriefingThe threat has evolved.
The stakes haven’t.
Your mission. Your customers. The people who depend on you.
AI is changing how these are attacked—and what it takes to protect them.
More convincing.
More coordinated.
Moving faster.
Identity can be imitated.
A familiar voice. A plausible invoice. A request that appears to come from the right person. Trust needs independent verification.
Intrusion can be accelerated.
AI-assisted operations can connect reconnaissance, stolen credentials, and data theft across organizations and their suppliers.
Your own AI creates new boundaries.
An agent with access can take consequential action. Its authority needs to be as carefully designed as its intelligence.
Published research. Measured claims.
2025 reported losses in complaints referencing AI
FBI IC3 · COMPLAINT DATA~200↗Downstream organizations in one supplier compromise
ANTHROPIC · REPORTED CASEA different kind
of defense.
Software that adapts. Expertise that connects the dots.
Built around the workflows where failure matters most.
One request.
A chain of consequences.
Follow an attack through a critical workflow.
See where trust is tested—and where a defense can intervene.
The familiar request.
A convincing identity is not the same as authority.
Source: Financial Crimes Enforcement Network“Please update the payment details.”
A fictional supplier sends an urgent bank-detail change, reinforced by a call that sounds familiar. The request looks routine.
A message and a voice are signals to examine. Neither independently authorizes a payment.
Understand. Build.
Verify. Adapt.
A practical starting point. A path to stronger defenses.
A process designed to evolve with your organization.
Understand
Map a critical workflow. Identify the decisions, identities, and dependencies that deserve attention.
Build
Develop the software, verification steps, and controls needed to address the findings.
Verify
Test agreed scenarios. Document what holds, what needs work, and what remains unknown.
Adapt
Revisit defenses as systems, business processes, and attacker capabilities change.
CONNECTED SYSTEMS. CONSEQUENCES THAT MATTER.
The attack follows connections. So do we.
From an identity to a supplier. From a request to a payment. From an AI tool to a consequential action. Build the defense around the whole workflow.
Different organizations.
Shared resolve.
The frontier is moving.
Stay grounded.
What the latest research actually says about AI-enabled attacks, fraud, and the defenses that still matter.
Read the launch briefingMeet the threat.
Change the outcome.
Start with your critical decisions and services.
Protect the customers and communities who rely on them.
