Friday, October 2, 2026

Lavender and Habsora - Israel's AI-Assisted Targeting Systems

 

In the war in Gaza that began in late 2023, two names appeared repeatedly in investigative reporting about how modern militaries generate targets: “Lavender” and “Habsora”. They are two distinct but complementary Artificial Intelligence systems used by the Israeli Defense Forces (IDF). One focuses on Who, the other on Where.

1. The Names

Both names are intentionally mundane. Habsora is Hebrew for "The Gospel" or "The Tidings". It is sometimes written as "The Gospel."

Lavender has no Hebrew meaning; it follows a pattern in the IDF where intelligence systems are given English flower or color names - Blue Wolf, Red Wolf, Wolf Pack.

2. Habsora / The Gospel: The Infrastructure Targeting System

Habsora was developed by Unit 8200 of the Israeli Intelligence Corps of the IDF. Habsora was introduced and first used in operations in 2021.

Its function is to identify physical structures, not people. The program is described as using vast amounts of data to identify, with the help of artificial intelligence, physical targets such as rocket launch centers, headquarters, or potential storage sites.

According to reports, Habsora is designed to mark buildings and structures that the IDF claims militants operate from. It is categorized into four types of targets: tactical targets, underground targets, power targets, and family homes.

What made Habsora notable was its scale. Investigations described it as an AI-powered targeting system that identifies and generates attacks on up to 100 targets a day, compared to about 50 targets a year produced by human analysts before the war. This was described by a former officer as a "mass assassination factory."

A critical feature attributed to Habsora is that it also includes an estimate for the number of civilians who may be killed in attacks on private residences that it recommends bombing in advance.

3. Lavender: The Individual Targeting System

The exact date of development of Lavender remains classified.

Whereas Habsora focuses on buildings, Lavender focuses on marking individuals to include on a kill list.

Lavender is described as an AI-enabled database that assigns risk scores to Gazans based on patterns in their personal data to identify suspected operatives.

How does it work:

Data ingestion: Lavender uses data from the mass surveillance of Gaza, where the entire telecommunications network is monitored, to identify targets based on their phone calls, movements and interactions.

Analysis: It works by analyzing information collected on almost all 2.3 million Palestinians in the Gaza Strip through a mass surveillance system. It assesses the likelihood of a person belonging to Hamas using a non-transparent ranking system.

Scoring: It uses machine learning to assign every resident a numerical score indicating their suspected likelihood of being affiliated with Hamas.

Output: Lavender generated a list of roughly 37,000 Palestinians as suspected operatives in the early weeks of the war. Other reports put the number as high as 37,000 men flagged as potential targets.

Lavender does not operate alone. Investigative reports from +972 Magazine and Local Call, which first broke the story in November 2023, said Lavender works alongside Habsora and a third system nicknamed "Where's Daddy?" — a system reportedly used to track when a flagged individual entered their family home to time a strike.

4. How The Two Systems Interact

Conceptually, the workflow described by sources is:

1. Lavender asks: Is this person a militant? It provides a probability score (1-100).

2. Habsora asks: Where is a legitimate military infrastructure target? It provides coordinates and target type.

3. Where's Daddy?[1]  asks: When is the person from step 1 inside a building from step 2?

4. Human operators are then presented with the fused recommendation.

The IDF system, known as Lavender, was used to find and target alleged militants. According to the Israeli news report, Lavender works alongside another Israeli system known as Gospel. The former targets individuals.

5. Official Response and Controversy

The IDF has responded to these reports, particularly after the documentary "NAZA" featured testimonies from intelligence officers about the systems.

The army stated that its Lavender AI targeting program is not a list of targets pre-approved for strike, but rather a database to help cross-reference intelligence sources. A person's inclusion in the Lavender database does not mean that individual is a target and does not authorize a strike against them.

The IDF categorically rejects allegations that it used AI to indiscriminately target civilians, saying that all strikes require independent human verification, compliance with international law, and proportionality assessment.

Critics, including human rights lawyers and AI ethicists, raise three concerns:

A. Automation bias: If a system is 90% accurate and it flags 37,000 people, 3,700 are misidentified.

B. Scale without oversight: Generating 100 targets a day changes the tempo of war beyond what traditional legal review was built for.

C. Dehumanization: Turning risk scoring into a list reduces complex human context to data points.

Proponents argue that AI reduces human error, can process far more intelligence to avoid misidentification, and that the systems are decision-support tools, not decision-makers. Every strike still requires a human commander.

6. Historical Significance

Lavender and Habsora are not science fiction weapons. They represent a shift from precision-guided munitions to “precision-guided intelligence” (PGI). For 2,500 years, lavender the plant was used from Egyptian mummification to Roman baths. Lavender the system represents a similar shift in utility — from a scent to a signal.

Together, Lavender and Habsora illustrate the central dilemma of algorithmic warfare: Can an algorithm that is statistically effective be legally and morally acceptable when its errors mean civilian death? And where does accountability lie — with the code, the officer who built it, or the officer who approved its recommendation?

That question is now being litigated not only in military courts, but in documentaries, academic journals, and international courts, making these two systems perhaps the most documented case study of AI in active conflict to date.

 

 

 

 

 

 

 



[1] Where is Daddy? is the third layer in the targeting chain. If Lavender answers “Who is a target?” and Habsora/the Gospel answers “What building is a target?”, “Where’s Daddy” answers “When is the person inside the building?” It is not a separate kill-list. It only works after Lavender has flagged someone. Its job is to follow that person until an operational window appears.

 

Thursday, October 1, 2026

HUMINT – Assessing the Veracity of the Information and Source


Human Intelligence or HUMINT is intelligence gathered by human sources and it remains irreplaceable in a world dominated by satellites, cyber tools and Artificial Intelligence. It is indispensable because it is the only form of intelligence that can answer "why", not just "what". It accesses intent and penetrates denial and deception.

HUMINT is often romanticized, but in reality, it’s the most error, prone intelligence discipline. Every intelligence service has learned this the hard way.

Assessing the veracity of intelligence from a human source - HUMINT - is one of the hardest disciplines in intelligence. Two separate things are under the scanner at once:

 1. The source - Is the asset truthful, accurate, and in a position to know?

2. The information - Is what an asset reporting plausible, accurate, and can be corroborated?

 In NATO and most Western agencies this is formalized as the Admiralty System. Every report gets a 2-part grade like B2 or F6.

 

I. Source Reliability - A to F

 The person (agent) is graded, not the report 

A - Completely Reliable: History of accurate reporting, proven access, no motive to deceive. Rarely given.

B - Usually Reliable: Mostly accurate in the past, minor errors due to perception not deception.

C - Fairly Reliable: Mixed record. Accurate sometimes, inaccurate sometimes.

D - Not Usually Reliable: History of exaggeration, poor memory, past fabrications.

E - Unreliable: Proven fabricator, hostile, under opponent control.

F - Cannot Be Judged: New source, no track record, no way to test yet. Most new walk-ins start here.

 

What determines this grade: 

1. Access: Does the source have direct, legitimate access to the information? First-hand observation is very different from second-hand hearsay from a colleague. Always establish: How did you obtain this? When? Where? Who else was present?

 2. Track Record: Has previous information from this source been corroborated? A source who gave you 10 verifiable truths earns higher reliability.

 3. Motivation - MICE +: Why are they talking to you?

Money: Are they being paid? Debt problems?

Ideology: True believer, disillusioned, revenge?

Compromise/Coercion: Blackmail, fear of prosecution, pressure from family?

Ego: Seeking importance, thrill, desire to be a spy?

No motivation = major red flag. Money and revenge are the highest risk for fabrication.

 4. Vulnerability to Control: Is the source under surveillance? Could they be a double agent or dangle? Signs: too perfect information, information that fits your existing gaps too conveniently, reluctance to take a polygraph, inconsistent personal story.

 

II. Information Credibility - 1 to 6

 You grade the information itself, independent of who said it.

 1 - Confirmed: Independently confirmed by other sources, e.g. SIGINT, IMINT, OSINT, another unrelated HUMINT source.

2 - Probably True: Logical, consistent with known behavior, not yet confirmed but no contradictions.

3 - Possibly True: Plausible, but not enough context to judge. Fits some known facts, conflicts with none.

4 - Doubtfully True: Plausible but conflicts with other reliable information, or lacks logic. Needs more work.

5 - Improbable: Contradicts established facts, defies logic or technical reality.

6 - Cannot Be Judged: No basis to evaluate, new topic with no other sources.

 

What determines this grade:

 1. Internal Consistency: Does the story hold together over repeated tellings? Fabricators often add detail on retelling. Truth-tellers often say "I don't remember" - which is actually a good sign.

Use detailed chronological debrief: Have them tell the story forward, then backward, then out of order. Fabricated stories collapse when told backward.

2. Plausibility and Technical Accuracy: Does the information violate physics, economics, or known enemy doctrine? A source claiming a new missile has 10,000km range when their industry can't produce the fuel is suspect. Use subject-matter experts to vet jargon and details.

3. Corroboration: The gold standard.

Does it match other INTs?

Does it match open sources?

Does another independent human source report the same thing without collusion?

Is it corroborated indirectly - e.g., source says Unit X is moving, and you see increased logistics traffic on satellite?

Beware circular reporting - Source A tells Source B who tells you, looks like two sources but is one.

4. Specificity vs. Vagueness: Truthful reporting is often messy, incomplete, with caveats. "I saw 3 T-72s at approximately 14:00 near the crossroads, I think, maybe 4" is more believable than "I saw exactly 10 T-72s of the 42nd Division at grid 123456 at 14:32 hours." Excessive precision is a classic sign of fabrication.

5. Counter-Deception Checks: Is the adversary trying to feed you something?

Is it too good to be true? Does it exactly answer your Priority Intelligence Requirements?

Chicken feed: Is the adversary giving you small true secrets to build credibility for a big lie later?

What does the adversary gain if you believe this?

 

III. Practical Handling Process

 A professional assessment follows this sequence:

1. Pre-Debriefing: Check source file. What has he said before? What is his assessed access? What does he expect in return?

2. Controlled Collection: Don't ask leading questions. Use open-ended: "Tell me what happened." Don't show him you already know something. Hold back some known facts as test questions.

3. Post-Debrief Checks:

Time-lag test: Re-interview after weeks. Does detail remain stable?

Lifestyle check: Is his spending matching his income?

Technical validation: Can his claims be tested?

4. Formal Evaluation: Assign Admiralty rating, e.g., C3 - Fairly reliable source, possibly true information. Write a source comment: "Source has indirect access via... No independent corroboration yet. Motivation is financial..."

No single report from a human source, even an A1 source, should drive policy alone. HUMINT is inherently the most prone to human bias, memory error, and manipulation. Veracity is established over time, through consistent, corroborated reporting, not through one brilliant scoop.

 

Key Analytical Steps for Credibility Assessment 

* Evaluate Motives & Leverage: Examine why the source is talking (e.g., financial gain, revenge, coercion, ideological alignment, or self-preservation). Hidden agendas heavily distort HUMINT. 

* Distinguish Primary vs. Secondary Access: Determine if the source witnessed the event firsthand (primary) or heard it through rumors, gossip, or "Chinese whispers" (secondary). 

* Check for Circular Reporting: Ensure the "independent confirmation" isn't just the same rumor echoing back through a different channel.

* Cross-Corroborate with Technical/Open Sources: Triangulate the HUMINT claims against OSINT (open-source intelligence), SIGINT (signals intelligence), or GEOINT (geospatial intelligence) to validate physical realities. 

* Account for Bias & Deception: Remember that human sources filter facts through personal blind spots, fears, or active state-sponsored disinformation (white, grey, or black psyops). 


The strange case of Yehuda Gil 

The classic case of how human source can fabricate intelligence is of Yehuda Gil / RED FALCON - the failure that almost started a war in 1996.

Who was he?

Yehuda Gil, 63 at the time of his arrest, had served 30 years in Mossad. Recruited in 1970, the Arabic speaker from Libya, veteran of the post-Munich assassinations and Operation Moses, instructor at the Mossad academy who taught a course called "lying as an art".

His source, recruited in 1975, was codenamed Red Falcon - described as a relative of a senior Syrian general with proximity to the Syrian military establishment.

For 15 years the operation produced credible intelligence and Gil's reputation became currency inside Mossad.

What did he fabricate?

He fabricated troop movement and war-intentions reports attributed to Red Falcon.

The most dangerous one:

On the morning of September 1st, 1996, the Israeli military moved troops to the Syrian border in preparation for a war

The report said: Syria was planning a lightning strike. Damascus intended to retake the Golan.

IDF Northern Command went to elevated readiness, commanders pushed to strike first. While the IDF went as far as moving troops toward the border, Israel's military intelligence service and Defense Minister Yitzhak Mordecai strongly questioned the validity of Gil's single-source report. They resisted full mobilization, avoiding a devastating war. Prime Minister Netanyahu had the order on his desk - he didn't sign. The alert was stood down.

It was not the first time. In summer 1981 Gil had filed a similar Red Falcon report about imminent Syrian attack that reached Mossad Director Yitzhak Hofi, who had lived through the 1973 surprise attack. That report triggered a push to call up all reserves.

Investigators later found: Gil had fabricated at least one meeting entirely and materially altered others - intelligence he filed was not intelligence, it was invention, his own knowledge, his own assumptions.

Why out of frustration?

The textbook HUMINT failure triangle: access loss + ego injury + financial incentive.

 1. Access collapsed: In 1989 Red Falcon retired from the Syrian army. Whatever access he had was gone - an older man removed from circuits where intelligence was generated. Real flow stops at retirement.

 2. Career frustration: Same year, Gil was passed over for promotion. The organization concluded he was a brilliant individual operator but lacked temperament for senior leadership. His ceiling was where he stood. He formally retired in early 1990s.

Losing the source meant losing the foundation of his standing.

 3. The workaround: Instead of closing the file, Gil came back. He told Mossad that Red Falcon had reactivated, would speak only to him, not a new handler.

Every verification attempt was blocked with a plausible excuse that traced back to the source's preferences:

When Mossad wanted a second officer to join, Red Falcon refused. When they wanted recordings, the equipment malfunctioned.

And the mechanism was simple: Mossad would brief Gil beforehand on what questions they needed answered, and Gil would file reports that answered them. The raw material was his own decades of expertise about Syrian doctrine, not source reporting.

Money intended for the source stayed with Gil - about $35,000 in cash found in a safe in his home, with total diverted funds estimated around $200,000.

How was veracity missed?

This is exactly the dilemma faced by any intelligence organization - source reliability overrode information credibility.

Halo effect: His early genuine record purchased latitude and substituted for scrutiny.

Confirming bias: His reports described a Syria that had not abandoned the military option, even as peace talks advanced. That fit the post-1973 trauma model the institution already believed. As the analysis noted: the reports were not surprising, they were confirming.

Handler-only access: No system for trusted insider threat. The structure requires trust - the handler is the only point of contact.

The warning had existed since 1981. Aman Syria desk chief Ayi Halakmi had written in the margins of Gil's report that it did not match collateral intelligence and blocked the reserve call-up. Those notes were filed and never actioned.

The case was only reopened in the mid-90s when Paris station chief Yehia Mart and other analysts ran independent checks on Gil's claims and found they did not hold up, leading to attorney-general approval for a formal investigation.

Gil was arrested in 1997, tried in secret, sentenced to 5 years.

Lesson for HUMINT assessment: 

A single-source report should not drive policy. 

The most dangerous fabricator is not a new source who lies badly. It is a celebrated handler who knows exactly what real looks like, knows what headquarters wants to hear, and has become unverifiable because his legend became the outer wall of the operation.

 

Lavender and Habsora - Israel's AI-Assisted Targeting Systems

  In the war in Gaza that began in late 2023, two names appeared repeatedly in investigative reporting about how modern militaries generate ...