John Wiseman

A detailed aviation sectional chart of Washington, D.C., centered on the Potomac River. The map is overlaid with data points titled "2025-03-01 TCAS RAs," represented by a cluster of five small red diamonds. These "Phantom" Traffic Collision Avoidance System (TCAS) Resolution Advisories (RAs) are concentrated directly over the Potomac River between the Key Bridge and Chain Bridge. The RAs are situated immediately to the west of Prohibited Area P-56B, which protects the U.S. Naval Observatory (indicated by a blue circular symbol and a "WARNING AVOID" label). Key landmarks nearby include Georgetown University, the Washington Monument, and the Pentagon. A distance scale at the bottom left indicates the RA cluster is about 2 miles from P-56B.

8/ Map showing the phantom TCAS Resolution Advisories relative to the location of P-56B and the Naval Observatory.

7/ Link to my thread on the phantom TCAS RAs from when it happened in March 2025. bsky.app/profile/lemo...

6/ I forgot a link to their phantom RA paper! Here it is: casd-lab.github.io/publications...

5/ I think the summary is, we already strongly suspected that the USSS was behind the spoofing, and these researchers did a good job of analyzing the attack. It's less clear why the Secret Service tested it next to DCA and (to me, at least) how effective spoofing Mode C is for counter drone work.

4/ They cite an FAA spectrum analysis that identified activation of United States Secret Service counter-UAS system in a US Department of Defense facility near the DCA approach path, and they say this supports the idea that transmissions likely originated near P-56B. bsky.app/profile/lemo...

A screenshot of a VFR aeronautical sectional chart covering Washington, D.C. and Northern Virginia. Prohibited Areas P-56A and P-56B are prominently outlined with blue hashed borders over the National Mall and the Naval Observatory, accompanied by a bold text warning: "WARNING AVOID PROHIBITED AREA." The chart shows the Potomac River and various landmarks including the Washington Monument, the Capitol Building, the Pentagon, and Ronald Reagan Washington National Airport (DCA). It is dense with technical aviation data, including altitudes, radio frequencies, and VFR transition routes like "Route 1."

3/ They analyze the source of the TCAS attack at DCA: "This analysis indicates that the probability of the transmitter being located within the P-56B area is 94.1% relative to the most probable location identified by our filter." P-56B is the prohibited airspace around the Naval Observatory.

2/ They said they'll present their analysis of the possible TCAS cyberattack at the Network and Distributed System Security (NDSS) Symposium, which starts Feb 23. casd-lab.github.io/en/publicati...

After discovering some dangerous vulnerabilities of the TCAS anti-collision system and successfully validating them in the laboratory, researchers from the University of Genoa and the CASD – Scuola Superiore Universitaria , with the support of the PNRR SERICS - Security and Rights in Cyber ​​Space partnership , have analyzed a real case: a series of TCAS anti-collision system alarms recorded over a 3-hour period by 10 airliners landing at Ronald Reagan Washington Airport on March 1, 2025 . The analysis indicates that the observed behavior is consistent with an in-flight cyberattack on the collision avoidance system, but in a previously undocumented variant . To protect future operational scenarios, the research team has also developed a new method for identifying the source of the attack .

1/ Last week the Italian researchers who have been looking at vulnerabilities in the TCAS system put out a press release about their analysis of the "phantom" resolution advisories that happened near KDCA last March. (Image via Google Translate.) life.unige.it/comunicati-s...

@obtusatum.bsky.social: "during January the largest increase in ICE detainees was at the Dilley Immigration Processing Center, now the only facility where ICE detains minors. There is currently a measles outbreak at this detention center, operated by the for-profit private prison company CoreCivic."

ProPublica collected handwritten letters from children held at the Dilley Immigration Processing Center.

Great article tracking local news stories about ICE detention centers. "After El Paso’s ME ruled migrant’s death a homicide, ICE sent the next body to an Army hospital", "Nicaraguan Man in ICE Custody Dies in Mississippi Hospital", "Lawsuit alleges inhumane conditions at ICE Adelanto facility"

I did a double take at that part. "Oh what's this now…"

This is well done.

Don't decrease the budget. Don't reform. Abolish it, and prosecute.

Immigrant whose skull was broken in eight places during ICE arrest says beating was unprovoked A Mexican immigrant whose skull was broken during his arrest by immigration officers last month in Minnesota says the beating was unprovoked.

ICE told nurses he “purposefully ran headfirst into a brick wall,” but they actually beat him with a steel baton. They caused 8 skull fractures and 5 brain hemorrhages.
apnews.com/article/immi...

Security is not my field, but "When we pointed Opus 4.6 at some of the most well-tested codebases (projects that have had fuzzers running against them for years, accumulating millions of hours of CPU time), Opus 4.6 found high-severity vulnerabilities, some that had gone undetected for decades."

Now I'm going down a rabbit hole of fixing up this 1993 checkers-playing program I wrote. It ran on a 16 MHz computer with 2 MB of RAM!

Opus 4.6's analysis

GPT 5.2 Codex's analysis.

2/ I have 3700 lines of C++ that plays checkers. I wrote it decades ago, but I did debug the hell out of it and ran it under valgrind an awful lot to try to find memory corruption issues. Opus 4.6 found 4 ways to crash it. Codex 5.2 found those 4 and 3 more.

1/ "Evaluating and mitigating the growing risk of LLM-discovered 0-days." Opus 4.6 has found more than 500 "high-severity vulnerabilities" in open source apps. Yikes. I ran it over some of my C++ code…
red.anthropic.com/2026/zero-da...

USAF bases drone inquiry ends without any suspects An MoD Police probe into sightings at USAF bases in Norfolk, Suffolk and Gloucestershire concludes.

"A major alert followed a US Air Force (USAF) announcement that its personnel had spotted unmanned aerial vehicles over the bases in Norfolk, Suffolk and Gloucestershire in November 2024."

"The investigation is concluded... no suspects were identified."

www.bbc.com/news/article...

Thank you for the coverage, @npr.org.

If you, dear reader, would like to help other people observing ICE in Minnesota, please contribute to our dash cam drive - more than a thousand have been distributed so far, and demand remains high: ottergoose.net/dashcam/

8/ Searchable.City is by Sean Hardesty Lewis, who's done some other interesting projects and writes them up extremely well. seanhardestylewis.com

7/ I do think if there was a cat in the aerial imagery around Edwards AFB that PIMINTO would find it.

A screenshot of the web application "PIMINTO" (Personal Imagery iNtelligence Operator). The interface is split into two main panels. On the left, a search sidebar shows a text query for "cats" with 2,688 filtered results and a list of ranked matches including small preview thumbnails. The right panel displays a large map centered on the California desert near Edwards Air Force Base. An aerial imagery overlay is covered in dense clusters of bright red and orange square data points, indicating locations where the search term was detected.

6/ PIMINTO doesn't do a good job of finding cats either, but because it uses vector similarity it can rank results. If its best match isn't actually a cat, you probably don't need to look at the other hundreds of results.

5/ Searchable City's semantic search of geospatially-situated imagery is very similar to my PIMINTO project, but PIMINTO uses vector similarity instead of caption generation. piminto.obliscence.com

A screenshot of a web application titled "Searchable.City" showing a dark-themed map of Manhattan, New York. The map is covered in thousands of bright yellow data points, forming a dense grid that follows the city's street layout. A search bar in the top-left corner contains the word "fire" and indicates 17,534 matches. Below the search bar, suggested filter tags include "graffiti," "hotel," "streetlight," "flower," and "tourist." The map interface includes standard navigation controls and a small "mapbox" logo in the corner.

4/ It didn't work to help me find cats. ...or did it?

A screenshot of a software interface displaying a Google Street View image of a helicopter flying through a cloudy sky above a city skyline. Below the image, the interface shows geographic coordinates (40.75425, -73.95694) and a heading of 135° SE. A section titled "DETECTED CONCEPTS" lists various AI-generated keyword tags including "flying," "cloudy," "photograph," "architectural," and "openness." It says there are 314 more tags.

3/ This technique of indexing captions generated by AI works well for some use cases, not as well for others. The AI determines what's salient in the image. This system seems to be tagging most photos (locations?) with about 300 tags.

Screenshot of a map of NYC. There's an inset street view image showing a helicopter flying.

2/ Here's an example where I searched for "helicopter".

Searchable.City Search NYC by what you can see. An open-vocabulary semantic atlas for querying urban elements like scaffolding, murals, awnings, trees, and more. Created by Sean Hardesty Lewis.

Really cool semantically searchable map of NYC. Looks like they ran thousands of Google Street View images through a visual language model to generate captions/tags, then indexed them. searchable.city

A line graph from METR titled "Time-horizon of software engineering tasks different LLMs can complete 80% of the time." The X-axis shows LLM release dates from 2020 to mid-2026, and the Y-axis shows task duration for humans, ranging from 0 to 1 hour. The data shows an exponential growth curve. From 2020 to early 2024, models like GPT-2, GPT-3.5, and GPT-4 remained near the 0-minute mark. Starting in late 2024, capabilities spike sharply: Claude 3.7 Sonnet is at roughly 10 minutes, Claude Opus 4 at 17 minutes, o3 at 25 minutes, and GPT-5 at 32 minutes. By mid-2026, GPT-5.2 (high) is shown reaching the 1-hour threshold. Reference tasks on the Y-axis include "Find fact on web" (10 min), "Implement a dictionary attack" (17 min), "Implement a simple webserver" (22 min), and "Train classifier" (50 min). Error bars on the most recent models indicate a range of uncertainty, but the trend line shows a vertical trajectory in AI task-handling capacity.

Here's the chart showing 80% success rate on METR, linear scale. The metric is amount of time a human would take to do the task, not the amount of time the model actually spends.