Microsoft’s New Security Tool Routes Most Work to a Smaller AI Model
Project Perception combines specialist agents, enterprise context and selective escalation to a larger model.
Microsoft has introduced a security system designed to do more than answer questions about threats. Project Perception is meant to connect signals from across an organization, reason about them with multiple AI agents, and pass approved actions into existing security tools.
The product enters public preview on August 3. Its first practical use is vulnerability management, where Microsoft is combining a new specialist model called MAI-Cyber-1-Flash with MDASH, the company’s existing system of more than 100 agents for finding, confirming and fixing software flaws.
The interesting part is not simply that another cybersecurity model exists. It is how Microsoft divides the work. The smaller specialist model handles up to 90 percent of tasks. The hardest 10 percent are escalated to GPT-5.4, a larger and more expensive model. That routing strategy could make continuous security analysis more affordable without asking one model to be equally good at every problem.
The announcement is easier to understand when its products are separated.
MAI-Cyber-1-Flash is the model. Microsoft says it is compact, optimized for code-heavy security work and derived from its MAI-Thinking-1 model.
MDASH is the operational harness around the model. A harness is the software that gives an AI model tools, assigns jobs, checks results and coordinates other agents. In this case, MDASH can use different models and specialized agents to reproduce a vulnerability, inspect a codebase and propose a repair.
Project Perception is the broader system. It brings together security signals, organizational context, models, agents and actions. Microsoft describes red-team agents that search for attack paths, blue-team agents that investigate risk and green-team agents that help remediate problems. It is intended to work across the company’s security products, including systems that manage identities, endpoints and sensitive data.
That distinction matters because the headline performance figure belongs to the configured MDASH system. It is not a score for MAI-Cyber acting alone.
Microsoft reports that MDASH with MAI-Cyber-1-Flash and selective GPT-5.4 escalation scored 95.95 percent on CyberGym. CyberGym is a benchmark built from 1,507 known vulnerabilities in 188 open-source projects. The task is not merely to describe a flaw. A system must inspect the relevant software and produce evidence that it can reproduce the vulnerability.
The company says the result is about 12 percentage points above the next system it compared and that the new configuration costs almost 50 percent less than its current MDASH setup. Those two claims point to the product’s practical proposition. A security team may be able to run more checks if routine work goes to a fast specialist and only ambiguous cases consume frontier-model capacity.
This is a familiar engineering pattern. Customer support systems route difficult cases to senior staff. Cloud platforms send different workloads to different processors. Project Perception applies the same logic to AI reasoning, with the added complication that errors can affect production systems.
Microsoft says the system includes role-based access, tenant isolation, encryption, audit records and sandboxed execution without internet access. These controls matter more than the model’s conversational fluency. A security agent needs permission boundaries, a record of what it did and a safe environment in which to test potentially harmful code.
The 95.95 percent result is substantial, but it remains a vendor-reported benchmark. Independent coverage noted that it had not appeared on CyberGym’s public leaderboard at the time of publication. Microsoft says the model was red-teamed and assessed by third parties, but it has not published enough detail for outsiders to reproduce every part of the system result.
CyberGym also tests known vulnerabilities in open-source projects. That is useful evidence for software analysis, but it does not establish how well Project Perception will handle a novel attack, a poorly documented internal application or a noisy corporate network. It says even less about whether an automatically proposed fix will preserve business logic.
The preview will therefore be a test of operations as much as intelligence. Buyers should ask which actions require human approval, how false positives are measured, what evidence accompanies each recommendation, and whether the system can explain why it escalated a case to the larger model. Pricing, regional availability and the exact set of supported integrations also remain unclear in the public announcement.
The immediate audience is enterprise application-security and security-operations teams, especially organizations already using Microsoft’s identity, endpoint and cloud-security products. The system is not being presented as a standalone model for general developers. Its value depends on the surrounding data, permissions, integrations and review process.
For those teams, the useful question is not whether AI can find vulnerabilities. Several tools already attempt that. It is whether model routing can deliver reliable findings at a cost low enough for frequent use, while keeping every proposed action governed and reviewable.
Project Perception offers a concrete architecture for that tradeoff. The public preview will show whether the controls and evidence are strong enough to turn an impressive lab score into a dependable workplace tool.
The Laser-Powered Drone Has Not Flown Yet
A new receiver powered a stationary propeller efficiently, but flight, tracking, range and safety tests still lie ahead.
A new receiver powered a stationary propeller efficiently, but flight, tracking, range and safety tests still lie ahead.
Several headlines this week suggested that researchers had charged a drone in mid-flight with a laser. The experiment was interesting, but it did not do that.
A team from the Civil Aviation University of China and Tsinghua University built a receiver that converted green laser light into electricity at 38.49 percent efficiency. The device was mounted beneath the wing of a stationary drone model and powered its propeller. The aircraft did not fly, and the next planned step is an outdoor test on a lightweight drone.
That gap between the headline and the demonstration does not make the research trivial. It shows where progress has actually occurred. The team improved the receiver and its cooling, two difficult parts of sending useful power through the air. The harder system-level problems remain open.
The receiver combines two energy-conversion devices. Its first layer is a perovskite laser cell, a semiconductor device tuned to turn concentrated laser light into electricity. Perovskite here refers to the material’s crystal structure. The researchers used cesium lead bromide, which responds well to blue and green light.
The second layer is a thermoelectric generator. It produces electricity from a temperature difference. Instead of treating all the receiver’s heat as waste, the tandem design uses some of that heat as another source of electrical output.
Heat is also the design’s main obstacle. Under a high-power laser, the researchers measured temperatures between 80 and 90 degrees Celsius. That level of heating reduces photovoltaic performance and can accelerate material degradation.
The team added antimony triselenide nanocrystals to the perovskite layer. These particles slowed heat flow and improved the material’s charge transport. The researchers also placed air channels in the model wing, allowing airflow from the propeller to cool the thermoelectric layer’s cold side. A larger temperature difference helps that layer generate more power.
Under 520-nanometre green light at an incident power density of 1.2 watts per square centimetre, the combined receiver reached the reported 38.49 percent power-conversion efficiency. That is a meaningful laboratory result for the receiving device.
It is not the efficiency of a complete airborne power system. Electricity must first run a laser on the ground. The beam then travels through air, where pointing errors, distance, haze and turbulence can reduce the energy that reaches the aircraft. The receiver converts only the surviving light, after which motors and electronics consume the resulting electricity. The published percentage covers one section of that chain.
The proof of concept showed that the receiver could produce enough electricity to turn a propeller on a stationary model. It did not demonstrate a closed loop that tracks a moving aircraft, maintains beam alignment, charges a battery and supplies the changing power needed for stable flight.
Those distinctions matter because a drone does not draw constant power. Takeoff, climbing, gust compensation and manoeuvring can produce sharp changes in demand. A receiver must remain illuminated while the aircraft changes position and attitude. Added hardware must also earn back its weight through longer endurance.
Outdoor laser transmission introduces another constraint. A concentrated beam must stop or redirect safely if it loses the intended receiver. Clouds, dust and atmospheric distortion may alter its path or intensity. The researchers explicitly identify real-time tracking and safe outdoor operation as work still to be done.
Independent coverage from Professional Engineering and Hackster correctly describes the setup as a stationary model. Live Science also notes in its image caption and later reporting that a real in-flight demonstration is planned rather than completed, despite using a stronger headline.
Battery endurance limits inspections, mapping, emergency monitoring and other drone work. Supplying energy remotely could extend missions without carrying a much larger battery. Laser power is attractive because a focused beam can deliver energy farther than the magnetic coupling used by a phone charging pad.
The new research advances that possibility by addressing a receiver problem that worsens as laser power rises. The tandem device recovers some waste heat, while the nanocrystals and wing airflow manage temperature. It also treats aircraft integration as part of the design rather than an afterthought.
The honest conclusion is not that unlimited drone flight has arrived. It is that one receiver crossed a useful efficiency threshold on a bench and powered an aircraft component while managing high heat. A flying demonstration would be the next newsworthy step. Reliable tracking, end-to-end efficiency, weather tolerance and beam safety will determine whether the idea becomes an operational tool.
How a Photonic Time Crystal Controls Light in Time
Researchers built a terahertz material whose optical response repeats fast enough to reshape how light behaves.
Researchers built a terahertz material whose optical response repeats fast enough to reshape how light behaves.
The phrase “time crystal” sounds like science fiction. In this case, it does not describe a machine that runs forever, nor the many-body time crystals studied in quantum physics. It describes a driven optical system whose properties repeat in time.
That distinction matters because a team from École Polytechnique, Collège de France and Helmholtz-Zentrum Dresden-Rossendorf has now demonstrated the first all-optical photonic time crystal. The result, published in Nature on July 29, moves an idea that had largely lived in theory and lower-frequency electrical systems into terahertz light.
Terahertz waves sit between microwaves and infrared light. They oscillate too quickly for ordinary electronics to control easily, yet the tools used for visible light do not always work well there. The new experiment shows one way to bridge that gap by making a material change in step with the wave itself.
An ordinary photonic crystal contains a repeating physical pattern. As light moves through it, the pattern allows some frequencies to propagate and blocks others. The principle resembles how a semiconductor creates allowed and forbidden energy bands for electrons, although here the engineered object is light.
A photonic time crystal moves the repetition from space into time. Instead of encountering alternating regions as it travels, the light experiences a material whose optical response is being switched back and forth at a regular rate.
This periodic change creates copies of the system’s optical modes at shifted frequencies. When those copies interact, they can form a momentum gap, a range in which the usual stable wave solutions are replaced by modes that shrink or grow over time. The growth is not free energy. The external drive supplies it, much as a child on a swing gains amplitude when pushed at the right rhythm.
The difficult part is timing. To produce the photonic time-crystal regime, the material must change strongly, coherently and on a timescale comparable to one cycle of the light. Earlier demonstrations relied mainly on electrical circuits at microwave frequencies. Optical materials have generally been too slow, too weakly tunable or too lossy to satisfy all three conditions together.
The researchers built a metamaterial, an engineered structure whose behavior comes from its geometry as well as its ingredients. It contains micrometre-scale gold cavities, a silicon nitride insulating layer and indium antimonide, a semiconductor.
Inside each cavity, light couples to a surface plasmon. A plasmon is a collective oscillation of mobile electrons that can trap electromagnetic energy near a surface. The cavity has a natural resonance, like a small bell tuned to a terahertz frequency.
The team drove the structure with a narrowband terahertz pulse from TELBE, a high-field source at the ELBE accelerator in Dresden. The field accelerated electrons in the indium antimonide. Because that semiconductor has an unusual electronic band structure, the electrons’ effective mass changed as their momentum oscillated.
Effective mass does not mean the electrons gained ordinary resting mass. It is a way of describing how they respond to forces inside a material. Changing it altered the cavity’s kinetic inductance, which is the part of its electromagnetic behavior associated with electron motion. The resonance then shifted and oscillated twice during each cycle of the driving field.
That is the central mechanism. One terahertz field periodically retuned the material fast enough for another probe pulse to experience a repeating optical environment in time.
The researchers measured the structure’s reflectivity and phase with sub-cycle timing. Their spectra showed two driven optical modes moving together and then merging at what physicists call an exceptional point. Beyond that threshold, one mode became less strongly damped. The measured linewidth narrowed by about 40 gigahertz, corresponding to a reduction of more than 50 percent in non-radiative plasmonic losses.
Reducing loss is useful because plasmons normally dissipate energy quickly. If the drive eventually supplies enough gain to overcome all losses, the same approach could support a plasmonic terahertz laser. The paper’s model places that possibility within experimental reach, but the team did not demonstrate such a laser.
Nor is this a ready-made optical processor. The experiment depends on a specialized accelerator-based terahertz source. It operates around 0.77 terahertz and uses a carefully fabricated sample. Above the tested field strength, incoherent effects such as impact ionization begin to interfere with the clean modulation. Integration, energy cost and operation at other frequencies remain open engineering problems.
The practical consequence today is therefore narrower and more valuable than the broad application claims. Researchers now have an experimentally tested way to engineer optical loss and frequency conversion in time, rather than only through a static material pattern. It gives photonics a new control surface.
Future devices may use that control for tunable terahertz sources, communications or sensing. The immediate achievement is simpler. Light was used to retune a material at the pace of light itself, and the resulting dynamics matched the defining behavior of a photonic time crystal.