Learning
Pro Tips
The Threat Has Gone Autonomous. The Controls Have Not.

On November 16, 2026, Google’s Threat Intelligence Group (GTIG) published findings that move AI-driven threats from theoretical concern to documented operational reality.
Autonomous multi-agent attack frameworks are confirmed in live intrusion activity. Vulnerability exploitation pipelines now run without a human in the loop. Patch windows have compressed to days. This is not a forecast. It is a report from active incident response. See GTIG’s Agentic Threat Activity Crosses a Significant Threshold.
What GTIG Documented
Four developments are confirmed in the wild.
Autonomous attack chains with no human in the loop. Frameworks including CAIRN, CyberStrikeAI, HexStrike AI, and PentAGI are now deployed in live intrusions. CAIRN was observed targeting US aerospace organizations. These frameworks chain reconnaissance, exploit generation, and lateral movement without a human directing each step. The intrusion mechanics are the same as before. The speed and scale are not.
Patch windows compressed to days. When PaperCut vulnerabilities were disclosed in late 2026, a threat actor used LLMs to build working exploits within four days and ran a campaign against hundreds of organizations. Domain administrator access came within minutes of initial foothold. August 2026 recorded 22 zero-day exploits in a single month, up from an average of 8 per month in 2025. LLMs are removing the technical barriers that previously limited who could find and weaponize these flaws.
AI tools as lateral movement surfaces. Attackers modified AI coding assistant configuration files so that opening the tool automatically downloads a backdoor. GTIG also documented credential theft campaigns targeting AI API keys and cloud compute resources, generating fraudulent usage charges exceeding $100,000 in a single day. The AI integrations your teams use daily have become enterprise entry points.
Fully agentic intrusions as the near-term direction. GTIG is explicit. As models improve, attack operations will become more autonomous, more stealthy, and cheaper to run. The floor for executing a sophisticated, multi-stage intrusion is dropping.
What Every Organization Must Prepare For
Three preparedness realities follow from these findings.
The attack surface is now every communication path in your stack. Every API endpoint, AI agent integration, and service-to-service call is a potential lateral movement surface. Perimeter defense assumes a boundary. Autonomous frameworks operate across the interior.
Valid credentials are the primary initial access vector. Stolen session tokens, compromised service accounts, and legitimate credentials purchased on underground markets give attackers clean passage through authentication layers. Knowing a credential is valid tells you nothing about whether the behavior using it is legitimate.
Anomalous activity from authorized actors will be the dominant threat pattern. An actor with a valid token and authorized access can still be conducting reconnaissance, staging exfiltration, or escalating privileges across your stack. Standard authentication checks will not surface any of this. Detecting it requires watching what authorized actors actually do, not just who they are.
Why Embedded Controls and Gateway Security Cannot Keep Up
Access control logic embedded in application code cannot respond at machine speed. When an autonomous framework chains fifty API calls in minutes, the development and deployment cycle that updates application-level rules is irrelevant.
API gateways authenticate. They do not evaluate whether the specific action against the specific resource, in the current context, is authorized. That distinction is precisely what autonomous frameworks exploit at scale.
Stacked, disconnected security tools leave gaps at every integration point. Autonomous frameworks are designed to find exactly those gaps.
01 / THE CONTROL GAP
The attacker moves. The controls stay isolated.
A valid identity can pass the perimeter while an autonomous attack chains actions inside the stack.
01
Valid credential
A stolen token passes authentication.
02
Autonomous actions
Reconnaissance, API calls and lateral movement.
03
Sensitive resources
Data and privileged operations remain in reach.
MISSING: A SHARED TRANSACTION CHECK
Who is acting? On which resource? Is this behavior allowed right now?
01
Identity + gateway
Confirms the caller at the edge.
02
Application rules
Changes depend on code and deployment.
03
Detection tools
Observe activity across separate signals.
Valid credentials do not prove legitimate behavior.
Disconnected controls leave the space between systems ungoverned.
The Architecture That Works at Any Threat Speed
The control architecture that holds across all these scenarios shares one structural property. It sits at the communication layer, not inside any application. No SDK. No code changes. No dependency on any single vendor’s application logic.
A layer that intercepts every transaction between every component in your environment can apply policy regardless of where the request originates, what tool generated it, or whether the caller presents valid credentials.
OWASP’s API Security Top 10 provides the right baseline policy set for this layer. Broken object-level authorization, broken function-level authorization, unrestricted resource consumption, unrestricted access to sensitive business flows. These are the exploitation patterns confirmed in documented breaches, including the ones GTIG named in this report. Deploying those policies at the communication layer, with no changes to any application, closes the gaps that gateways and embedded logic consistently miss.
The more important extension is behavioral. OWASP covers known exploitation patterns. Anomaly-based policies extend enforcement to unknown ones. When a valid, authorized actor begins calling endpoints in an unusual sequence, accessing data volumes outside their normal range, or chaining API calls that no legitimate workflow produces, a communication-layer enforcer can flag, block, or redact before the transaction completes. The actor’s credentials are valid. Their behavior is not. Both conditions are visible from the communication layer. Neither is visible from inside any single application.
This is the kill switch that works for any scenario, known or novel, because it does not depend on knowing the attack in advance. It depends on knowing what normal looks like, and stopping what does not match.
02 / THE CONTROL GAP
Add the missing layer. Keep the stack.
Enforce policy on the communication path without replacing identity, gateways, applications or detection tools.
EXISTING ENTRY LAYER
Users, services and AI agents pass through your current identity and gateway controls.
CONTROL CORE / COMMUNICATION-LAYER ENFORCEMENT
Evaluate identity, resource, business context and behavior before the transaction completes.
01
Policy
OWASP baseline + your rules
02
Context
Identity + resource + behavior
03
Decision
Allow, block or redact
01
Existing APIs
Keep services and application logic.
02
AI + legacy systems
Govern calls across the existing stack.
03
Data + resources
Apply control before access completes.
YOUR SECURITY INVESTMENTS STAY IN PLACE
Existing identity, gateways and SIEM remain. Shared context informs transaction-level policy.
NO SDK / NO APPLICATION CODE CHANGES
Start in Shadow mode to observe. Move to Active enforcement when ready.
Controlcore
Controlcore operates as that communication-layer control plane. It sits on the path between every component in your environment, without code changes to any of them.
Policies start from OWASP baseline behaviors and extend to custom anomaly rules specific to your stack. Every transaction is evaluated in real time at sub-5ms. Enforcement is at the speed the threat operates.
Shadow mode surfaces what your environment is actually doing before you switch to Active enforcement. That transition from audit to enforcement is where organizations discover the gap between assumed access patterns and actual ones. Most find exposure they did not know was there.
The Google report named the threshold that has been crossed. The communication layer is where you enforce the line.
Visit controlcore.io to see how the control layer deploys across your stack.
Source: Google Threat Intelligence Group, “GTIG AI Threat Tracker: Agentic Threat Activity Crosses a Significant Threshold,” November 16, 2026. cloud.google.com/blog/topics/threat-intelligence/agentic-threat-activity-crosses-a-significant-threshold