X-62 VISTA Completes 27 AI-Controlled Intercepts Using Live Sensor Data

X-62 VISTA completed 27 AI-controlled intercepts across eight flights, giving google news readers an eye-catching number and a much more consequential technical shift. An AI agent reportedly used live infrared sensor data to locate a crewed T-38 and maneuver the experimental fighter toward an intercept position. The software was no longer flying toward a target supplied through a simulation.

That distinction changes the story. Earlier X-62 experiments established that machine-learning software could maneuver a fighter against simulated or crewed opponents under controlled conditions. The new HAVE HEAT tests connected sensing, interpretation, decision-making, and aircraft control during live flights.

The result is not an autonomous combat aircraft ready for deployment. Lockheed Martin and the U.S. Air Force still have not published the performance data needed to judge reliability under operational pressure. However, the experiment moves the central question from whether AI can fly tactical maneuvers to whether it can make useful decisions from imperfect sensor inputs.

That question matters well beyond one modified F-16. The Air Force is developing Collaborative Combat Aircraft, or CCAs, which are uncrewed aircraft intended to operate with crewed fighters. Those systems will need to interpret sensor feeds and act within human-defined limits without waiting for constant control inputs.

The X-62 test therefore pressures every autonomy program that remains dependent on clean simulation data or tightly scripted demonstrations. A convincing flight maneuver is valuable. Closing the full loop from sensor detection to physical action is the harder test.

What Changed During the X-62 VISTA Flights

The important result was not simply that AI flew a fighter, but that it acted on live targeting data from an operational sensor.

The U.S. Air Force Test Pilot School conducted the HAVE HEAT campaign at Edwards Air Force Base during April 2026. Lockheed Martin disclosed the results on August 4, alongside an Air Force account of HAVE HEAT and a parallel integration effort called HAVE HOLIDAYS.

According to the company’s intercept release, the X-62 performed 27 AI-controlled intercepts during eight flights. The target was a live T-38 Talon, a two-seat supersonic trainer used by the Air Force.

The X-62 carried Lockheed Martin’s Legion Pod, an infrared search-and-track sensor known as IRST. Unlike radar, an IRST system detects heat signatures without transmitting radio-frequency energy that could reveal the searching aircraft’s position.

The pod generated real-time information about the T-38. That information entered the aircraft’s autonomy architecture, where an AI agent interpreted the feed and controlled the X-62’s flight path. The aircraft then maneuvered into what Lockheed described as a tactical intercept position.

This was a closed-loop test. In a closed loop, the system observes its environment, decides how to respond, takes action, and uses new observations to adjust its behavior. A target’s changing location can therefore alter the aircraft’s next maneuver without a person manually entering every command.

The Air Force account confirms that AI agents ingested live infrared data and directed the aircraft toward a target in real time. It does not provide the 27-intercept figure, which comes from Lockheed Martin.

That attribution matters. Google news headlines can compress a company-reported test result into an apparently settled performance record. The available public material confirms the campaign and its sensor-to-aircraft design, but leaves several evaluation details undisclosed.

Neither organization has released the geometry of each intercept, the starting conditions, the success threshold, or the rate of human intervention. They also have not disclosed how often the sensor lost its track or how the agent responded to uncertain data.

The tests nevertheless advanced beyond the X-62’s earlier public demonstrations. During the 2023 Air Combat Evolution campaign, the aircraft conducted within-visual-range engagements against crewed F-16s. Human operators provided target information while the autonomy system controlled the flying.

In HAVE HEAT, the AI received target information from the aircraft’s own sensor integration. Chase Kohler, a spokesperson for the 412th Test Wing, told Air & Space Forces Magazine that the AI locked onto targets itself during the newer campaign.

That shift connects two areas often tested separately. Sensor engineers focus on detection and tracking, while autonomy teams focus on aircraft behavior. HAVE HEAT placed their outputs in one active chain.

The X-62 remained a supervised test aircraft throughout the campaign. It is a heavily modified two-seat F-16D, and safety pilots can disengage experimental control. That human safety layer allows developers to test lower-maturity software without treating every unexpected behavior as a probable aircraft loss.

The number 27 shows repetition across multiple flights, which is more informative than a single choreographed run. Yet repetition alone does not establish combat reliability. The deeper value lies in demonstrating that the complete architecture functioned often enough to support continued testing.

Why Google News Headlines Undersell the Sensor-to-Action Test

A fighter that follows AI-generated steering commands is one problem; a fighter that decides where to fly from live sensor evidence is another.

The distinction begins with data quality. Simulated target information can arrive in a consistent format with precise coordinates and predictable timing. Real sensors produce measurements affected by range, angle, atmosphere, background heat, aircraft motion, and other sources of uncertainty.

An autonomy agent must translate those measurements into a useful tactical picture. It must then choose a maneuver that respects the aircraft’s performance limits, test boundaries, and safety constraints. Each added step creates another place where errors can compound.

That is why the sensor-to-action loop is the campaign’s central mechanism. It joins perception and control inside a moving aircraft rather than treating them as separate demonstrations. Lockheed says its agents consumed classified infrared feeds and executed maneuvers in real time.

The company has not disclosed the underlying model design or training data. It calls its agent-generation capability Supermassive and says integration and ground testing took three months. Public information does not show how much of the system relies on machine learning versus conventional control software.

That mixed architecture would be normal. An autonomous aircraft does not need one model to handle every task. Engineers can assign perception, tactical selection, flight control, and safety enforcement to different components.

The X-62 is particularly useful because experimental software does not directly replace the aircraft’s entire control system. Its System for Autonomous Control of Simulation lets external autonomy software command the test environment. Independent protections can return control to the safety crew when necessary.

Those protections address a practical barrier that has slowed flight testing. Developers can revise software quickly, but traditional airworthiness processes assume aircraft configurations change less often. A protected testbed helps teams evaluate new agents without certifying each one as a complete operational system.

The Air Force Research Laboratory made that advantage visible in 2022. During an earlier campaign, teams switched autonomy algorithms aboard the X-62 within minutes and flew different experiments within hours. AFRL said VISTA accelerated full-scale autonomy testing by at least one year.

HAVE HEAT added an operational sensor to that rapid experimentation model. This makes the platform less like an airborne joystick for an AI pilot and more like a laboratory for integrated mission behavior.

The difference is relevant to future CCAs. An uncrewed aircraft flying alongside an F-35 or another command platform cannot depend on perfectly curated target coordinates. Communications can be delayed, interrupted, jammed, or constrained by emission-control requirements.

Onboard sensors provide another source of information. However, they also force the aircraft to distinguish ambiguous signals and decide when its confidence supports action. The software needs to respond appropriately when evidence changes or disappears.

HAVE HEAT does not establish that the X-62 solved those problems. It shows that one sensor, one live target type, and one experimental agent could operate together in a controlled campaign.

That narrower conclusion is still meaningful. Military autonomy programs often produce separate milestones for sensing, networking, and vehicle control. Integrating them can reveal timing mismatches and interface problems that remain invisible in isolated tests.

The campaign also illustrates why open mission architectures matter. A modular architecture defines interfaces that allow sensors, processors, and autonomy agents to connect without rebuilding the entire aircraft around each component.

The Air Force’s parallel HAVE HOLIDAYS program tested that proposition more directly. Engineers integrated a third-party autonomy agent, enhanced safety rules, and additional processing technology into the X-62’s Enterprise Open Mission System Architecture computer.

This supporting effort did not generate the headline number appearing across google news. It may prove equally important because operational autonomy cannot depend on one contractor’s tightly coupled hardware and software stack.

The Real Contest Is Integrated Autonomy Versus Scripted Demonstrations

The primary contest is not AI against a human pilot; it is integrated autonomy against demonstrations that avoid the hardest interfaces.

Dogfight imagery invites a simple comparison between human and machine performance. That framing can obscure what the Air Force actually needs from autonomous teammates. A CCA does not have to defeat every fighter in a close-range contest to provide operational value.

It may instead carry sensors, extend a formation’s reach, complicate an adversary’s targeting, or provide additional weapons capacity. Those missions require coordination, navigation, sensing, and compliance with human intent. They do not necessarily require an independent robotic ace.

The X-62’s 2023 campaign used dogfighting as a stress test. The Air Force reported 21 test flights, more than 100,000 lines of flight-critical software changes, and engagements against crewed F-16s. The aircraft approached within 2,000 feet during high-speed scenarios.

Safety pilots remained onboard and could disengage the AI. The Air Force said they did not activate the safety switch during those dogfights. That record supported the claim that nondeterministic software could be tested safely within a protected environment.

HAVE HEAT attacks a different bottleneck. The agent had to turn live sensor information into an aircraft response. This is closer to the workflow an autonomous teammate would encounter during a mission, even though the test remained controlled.

The pressure now falls on programs that demonstrate sophisticated maneuvers while supplying the autonomy system with preprocessed target data. Such tests can validate control behavior, but they leave perception and integration largely outside the evaluation.

The new campaign does not make previous work obsolete. Layered testing remains essential because teams need to isolate failures before combining systems. However, programs eventually have to place those layers together under realistic timing and data constraints.

The Air Force is pursuing that broader goal through several efforts. Project VENOM is modifying six F-16s to support autonomy experiments with a larger test fleet. X-62 remains a unique platform, but VENOM can generate more sorties and involve additional operators.

At the same time, General Atomics and Anduril are developing CCA prototypes for the service. These aircraft represent a different stage of the autonomy problem. They are purpose-built uncrewed platforms rather than crewed testbeds carrying safety pilots.

That difference creates a tradeoff. X-62 can test aggressive software changes because a trained crew and independent controls provide recovery options. An uncrewed prototype removes risk to onboard personnel but offers fewer ways to rescue an aircraft after a severe software failure.

The two approaches should therefore complement each other. X-62 can expose immature agents to flight conditions, while purpose-built CCA prototypes can test the actual vehicle, communications, and maintenance model intended for service.

International programs create additional competitive pressure. Boeing Australia’s MQ-28 Ghost Bat has become a prominent collaborative aircraft test platform, while several European programs are exploring autonomous teammates and remote carriers.

These competitors do not need to reproduce the X-62 exactly. They need credible ways to test how software interacts with sensors, networks, weapons, and human command structures. The strongest program will be the one that turns integration lessons into repeatable operational behavior.

The Air Force plans to begin fielding its first operational CCAs by 2030, according to reporting from The War Zone. That schedule leaves limited room for autonomy that works only in simulations or carefully staged flights.

The HAVE HEAT results support Lockheed Martin’s position as an important integration partner. They do not decide which autonomy supplier, sensor package, or aircraft design will dominate future deployments.

HAVE HOLIDAYS also complicates any vendor-centered interpretation. The Air Force integrated an agent from an unidentified non-prime contractor during that campaign. This suggests the service wants an environment where different suppliers can compete within shared interfaces.

That modular approach creates pressure on Lockheed too. The company benefits from its long relationship with X-62 and its role in the Legion Pod test. Still, open integration can make it easier for another autonomy developer to replace one component without replacing the full architecture.

The contest is therefore broader than Lockheed versus another defense contractor. It is a contest between systems that can absorb changing sensors and agents, and systems that perform well only as fixed demonstrations.

What the 27 AI-Controlled Intercepts Do Not Prove

The test establishes a credible integration milestone, but it does not establish combat readiness, independent target judgment, or dependable operation against an adversary.

The biggest missing information concerns the intercept standard. Lockheed says the aircraft reached a tactical position, but its release does not define the required distance, angle, timing, or persistence. Without that definition, readers cannot compare the 27 events with another test campaign.

The denominator also matters. The company reports 27 successful intercepts across eight flights, but it does not state how many attempts were planned. It is unclear whether every initiated intercept succeeded or whether unsuccessful runs were excluded.

Public reporting does not identify the variability across those events. The T-38 may have followed repeated profiles, or each attempt may have introduced different geometry and behavior. Those conditions determine how much generalization the agent displayed.

The sensor environment is another unknown. IRST systems are passive, which offers tactical advantages, but their observations can be affected by weather and background conditions. The Mojave test environment does not represent every climate or contested airspace.

The campaign also did not publicly demonstrate adversarial electronic warfare against the sensor-to-action chain. An opponent could attempt to disrupt communications, manipulate other data sources, or force the aircraft to reconcile conflicting tracks.

An infrared sensor does not make the system immune to deception. Countermeasures, terrain, cloud cover, and multiple nearby heat sources can complicate tracking. The released material does not explain how the agent handled ambiguous or lost tracks.

Human involvement requires careful language too. The aircraft autonomously maneuvered after consuming sensor data, according to the official accounts. That does not mean the AI independently selected an unknown aircraft as a target or received authority to employ a weapon.

Test staff defined the scenario, operating limits, and expected target. Safety pilots were aboard the X-62. Ground personnel also monitored the campaign, and the public release does not detail any corrections or resets between intercepts.

This distinction separates vehicle autonomy from lethal decision authority. An AI system can navigate toward an assigned target while humans retain control over identification, mission authorization, and weapons release.

The Air Force has emphasized human-machine teaming in its public descriptions of the X-62. The testbed’s value comes partly from preserving human supervision while researchers expose software to real flight conditions.

Reliability remains another open issue. Lockheed says its AI effectively and reliably closed the sensor-to-action loop, but that is a company assessment. No independent evaluation report or statistical reliability measure accompanied the announcement.

Twenty-seven intercepts can uncover integration problems and support a development decision. They are too few to characterize performance across the vast range of conditions a combat aircraft could encounter.

The safety architecture can also hide weaknesses that would matter on an uncrewed aircraft. Protective limits may prevent dangerous maneuvers, but they could also stop an agent from completing a mission when conditions fall outside its training.

That is not a criticism of the test design. Safety boundaries are necessary for experimental flight. The limitation appears when a protected research result is presented as direct evidence of operational capability.

The google news framing risks making that leap. “AI-controlled intercepts” sounds like an autonomous fighter independently found, classified, pursued, and defeated an opponent. The documented event was more specific and more constrained.

A live T-38 supplied the physical target. The Legion Pod generated infrared tracking information. An AI agent used that information to control a modified F-16 toward an intercept position, under test supervision and onboard safety protection.

That description is less cinematic. It is also technically more informative because it identifies the exact interface being evaluated.

The program will need harder tests before broad claims become justified. These should include multiple targets, incomplete tracks, sensor disagreements, changing mission priorities, degraded communications, and agents from different suppliers.

Evaluators will also need measures that go beyond successful intercept counts. Useful metrics include false-track responses, time to recover from lost data, safety-boundary violations, operator workload, and performance after unexpected scenario changes.

Transparency will remain limited because the sensor feeds and tactical methods are classified. That makes independent assessment difficult and increases the importance of carefully defined public metrics.

Readers should therefore treat the 27-intercept figure as evidence of repeated system integration, not a score proving that AI now outflies human combat pilots.

The Next Three Signals Will Decide Whether the Shift Holds

The decisive evidence will come from broader sensor integration, independent agent testing, and transfer from X-62 into operationally representative CCA programs.

The first signal is the planned installation of RTX’s PhantomStrike radar. The compact active electronically scanned array will give X-62 a sensor with capabilities and constraints different from the Legion Pod.

Air & Space Forces Magazine reports that installation, integration, ground testing, and limited flight tests are expected by the end of 2026. More involved testing is expected in early 2027.

Radar integration would strengthen the article’s central judgment if AI agents can combine its data with infrared tracks during flight. Multiple sensor types can complement each other, but they can also produce conflicting estimates that require careful reconciliation.

A test using radar alone would still extend the architecture. A test fusing radar and infrared information would be stronger because it would challenge the agent to reason across different data streams.

Failure to complete the upgrade on schedule would not invalidate HAVE HEAT. It would show that moving from one integrated sensor demonstration to a richer mission system remains difficult.

The second signal is whether the Air Force flies third-party agents under comparable conditions. HAVE HOLIDAYS integrated a non-prime autonomous agent and tested enhanced rules designed to keep autonomous vehicles within user-defined limits.

The public account does not say that this third-party agent performed the same live intercept mission. A future campaign that applies consistent evaluation criteria to agents from several vendors would test whether the architecture is genuinely modular.

That result would also clarify where the value resides. If multiple agents can use the same sensor and aircraft interfaces, the X-62 architecture becomes a reusable test environment. If every new agent requires extensive custom integration, the promised speed advantage narrows.

The Air Force Test Pilot School plans to convene the first X-62 Research Symposium at Edwards from August 11 through August 13. That meeting offers an early opportunity for government, industry, and academic participants to define future experiments.

Readers should watch for specific commitments rather than broad autonomy language. Useful announcements would identify new sensors, external agent suppliers, safety evaluations, or common metrics for comparing test outcomes.

The third signal is transfer into VENOM and CCA flight programs. X-62 is a research accelerator, not the aircraft that the Air Force plans to deploy.

A lesson counts operationally when it changes how another platform senses, communicates, or behaves. Software may not transfer directly because each aircraft has different flight characteristics and mission systems. Interfaces, test methods, and safety evidence can still carry over.

The previous X-62 campaign showed that the testbed could support rapid algorithm changes and human-versus-AI engagements. HAVE HEAT extended the chain into live onboard sensing.

The next stage must show that those lessons reduce risk for an uncrewed aircraft operating with a crewed formation. Evidence could include a CCA responding to onboard tracks, coordinating with another aircraft, or continuing safely after a communications interruption.

That would strengthen the argument that sensor-driven autonomy is becoming an operational capability rather than remaining an X-plane specialty. Repeated delays or highly scripted CCA demonstrations would weaken it.

The 27 intercepts should therefore be read as a bridge, not a destination. They connect the AI pilot demonstrations of 2023 with the integrated, networked aircraft the Air Force wants later this decade.

For developers, the lesson is that model performance cannot be separated from interfaces, timing, and safety controls. For enterprise technology buyers, the pattern is familiar: an AI system creates value only when it can act responsibly on live organizational data.

For aviation readers arriving through google news, the best question is not whether the X-62 defeated a human pilot. It is whether future tests preserve this complete sensor-to-action chain while adding uncertainty, competing inputs, and fewer safety assumptions.

Watch the PhantomStrike integration, comparable flights by third-party agents, and direct transfer into VENOM or CCA tests. Those three signals will show whether HAVE HEAT opened a repeatable development path or produced one carefully bounded success.