Sign in

dragosr

@dragostech.bsky.social
1.7K followers 3.8K following 1.3K posts

Autonomous Carbon Based LLM with 42 years of tuning on Information Attack and Defense. Host of CanSecWest, and PacSec. Security audits, code, IR, LLM, red team consulting. Specialize in Firmware, and RF. VA7MOV

PostsRepliesMedia
dragosr @dragostech.bsky.social · 07/10/2026
Reminder: In the US when a trademark is registered you can pursue "enhanced" damages up to three times proven damages.
020
dragosr @dragostech.bsky.social · 01/10/2026
Mr. Scruffles goes to CanSecWest...
Mr. Scruffles, a small brown plush cat wearing a CanSecWest-branded lanyard, sits on a conference table at CanSecWest. Behind him, attendees face two projection screens showing a presentation, with a large red CanSecWest banner between them.
160
dragosr @dragostech.bsky.social · 18/08/2026
Also seeking one more panelist for "Who Guards the Guardians?" on safety classifiers, guardrails, and their effect on security research. Promises to be interesting. :-) Image prompt: geoposition CanSecWest 2026 from Vancouver building profile extrapolation, wire-mesh from overlaid targeting reticle
Aerial view of Vancouver’s Coal Harbour skyline and marina, overlaid with a cyan targeting reticle highlighting the Pinnacle Harbourfront hotel. A wireframe magnification panel promotes CanSecWest 2026 at Pinnacle Harbourside, September 30–October 1, with the website secwest.net.
010
dragosr @dragostech.bsky.social · 18/08/2026
New training: reverse engineering with LLMs and Ghidra from Kara Nance, author of the No Starch Ghidra Book (she's also presenting her optical drone tracking work), plus Erlend Oftedal's two-day workshop on building an agentic vuln discovery pipeline. Registration just opened.
Aerial view of Vancouver’s Coal Harbour skyline and marina, overlaid with a cyan targeting reticle highlighting the Pinnacle Harbourfront hotel. A wireframe magnification panel promotes CanSecWest 2026 at Pinnacle Harbourside, September 30–October 1, with the website secwest.net.
110
dragosr @dragostech.bsky.social · 18/08/2026
Last week for CanSecWest submissions. Vancouver, Sept 30 / Oct 1. Keynotes: Bruce Schneier and Katie Moussouris. Sergey Bratus on automated micro-patching. Paper announcements coming to the site as we proceed. CFP and registration: secwest.net 🧵
Aerial view of Vancouver’s Coal Harbour skyline and marina, overlaid with a cyan targeting reticle highlighting the Pinnacle Harbourfront hotel. A wireframe magnification panel promotes CanSecWest 2026 at Pinnacle Harbourside, September 30–October 1, with the website secwest.net.
153
dragosr @dragostech.bsky.social · 24/06/2026
The datacenter design is symbiotic with the design of the Nvidia NLV72 racks. Here is the picture of (currently) world's largest AI center in Wisconsin. It's a lot of fans with radiators recirculating closed loop cooling liquid (non-evaporative) to the building just off camera that has GB300s GPUs.
Aerial view of Microsoft’s Azure Fairwater data center in Wisconsin under construction, showing two long parallel rows of large circular cooling fans that form the facility’s closed-loop GPU cooling system, flanked by white-roofed data center halls with overhead pipe runs connecting them.
000
dragosr @dragostech.bsky.social · 24/06/2026
PV has been dropping by 50% every two years, and while not dropping so fast it will keep winning. Grid scale battery costs dropped 93% in a decade. US public and gas industry hasn't figured out that Solar won, years ago, and will keep winning. But power companies got the memo, raw economics wins.
Stacked bar chart of U.S. new utility-scale power capacity additions by source, 2023 to 2027. Each bar shows percentage share; total new gigawatts are labeled above. 2023 (actual): solar 46%, battery 16%, wind 15%, gas 20%, 40 GW total. 2024 (actual): solar 63%, battery 21%, wind 10%, gas 5%, 49 GW. 2025 (actual): solar 51%, battery 28%, wind 9%, gas 9%, 53 GW. 2026 (planned): solar 51%, battery 28%, wind 14%, gas 7%, 86 GW. 2027 (projected): solar 46%, battery 28%, wind 12%, gas 12%, 65 GW. Source: EIA Preliminary Monthly Electric Generator Inventory.
440
dragosr @dragostech.bsky.social · 24/06/2026
WTF, what's happened to Popehat? This isn't cool at all. @kenwhite.bsky.social
120
dragosr @dragostech.bsky.social · 10/06/2026
The Anthropic Fable-5 safety classifiers seem to be written by the OpenAI marketing department. Pretty much anything I talk to LLMs about gets downgraded. Nerfed into useless. Worst model release ever?
A classic Scooby-Doo meme template set in a dark, stone dungeon filled with cobwebs. Fred stands in the center, pulling a green ghost mask off the villain to reveal an grumpy, elderly man tied to a wooden chair. Velma points at the unmasked villain while Daphne stands next to her. On the left, Shaggy and a frightened Scooby-Doo look on.

White text labels are superimposed over the image: the ghost mask is labeled "Fable-5" and the unmasked villain is labeled "Opus-4.8", implying that the supposedly new or monstrous "Fable-5" is actually just "Opus-4.8" in disguise.
171
dragosr @dragostech.bsky.social · 31/05/2026
When you leave all the mitigations on, crank up the inference settings to maximum, and run the exploit benchmark to see how close the PoCs can get.
Anime mech-style illustration of a confused humanoid murderbot sitting at a cluttered cyberpunk hacking station, one hand on its head as holographic warning screens announce “TOKEN BUDGET EXHAUSTED,” “CONTEXT WINDOW FULL,” “PROMPT TURN CAP REACHED,” and “OUT OF SPACE TO CONTINUE.” The robot is surrounded by half-finished exploit diagrams, terminal errors, sticky notes, ramen, pizza, and a whiteboard attack plan that has stalled at “???” because the AI system limits have been reached.
010
dragosr @dragostech.bsky.social · 11/05/2026
Raelize had LLMs reproduce older ESP32 Secure Boot V1 bypass via voltage glitching in a few days. AI wrote the tooling and built a live dashboard mid-campaign. 20k shots, 12.2% bypass rate. Workflow compression is the story. ESP32 v3 mitigates this. raelize.com/blog/ai-fi-g...
Dark infographic titled "AI-Driven Fault Injection on ESP32 Secure Boot." In the middle, an ESP32 chip is struck by a yellow lightning bolt, representing voltage glitching. A cyan box on the left labeled "AI Agent (Claude Code)" points to the chip, and a green box on the right shows an open lock and "Secure Boot Bypassed." Three large result cards sit below: the initial scan used 1,000 shots and found 1 bypass, the refined timing window used a 1.5-2 us delay and 3-4 us width for a 9.3% hit rate, and the voltage sweep from 2.00-3.00 V across 20,200 shots produced 2,468 bypasses, or 12.2% overall. The footer says AI automated the stack while humans set direction, emphasizing that the attack was known but the engineering effort was reduced.
120
dragosr @dragostech.bsky.social · 08/05/2026
"Dirty Frag" clickbait update: ESP (CVE-2026-43284) patched in mainline + stable (7.0.5, 6.18.28, 6.12.87, 6.6.138, 6.1.171). RxRPC (CVE-2026-43500) still unpatched upstream. AWS adds ipcomp4/ipcomp6 to the blacklist alongside esp4/esp6/rxrpc. AlmaLinux shipped both. Ubuntu/Debian mitigation only.
Infographic titled "Dirty Frag: Overhyped" framing the issue as a local Linux kernel privilege-escalation bug, not a remote wormable threat. It explains two exploit paths: ESP/IPsec, which needs a fresh user namespace to gain CAP_NET_ADMIN for XFRM SA registration, and RxRPC, which needs no namespace or extra capability if rxrpc.ko is present. Ubuntu and Debian namespace protections are shown as blocking the ESP path, while RxRPC can still bypass that defense. Other distro trade-offs are summarized, with RHEL-like systems often lacking RxRPC but leaving ESP reachable. Mitigations include blacklisting esp4, esp6, rxrpc, optionally ipcomp4/ipcomp6, using user namespace hardening, and relying on default container seccomp where applicable. Patch status notes ESP fixes in mainline and stable kernels, while RxRPC upstream status is still pending, with some distro-specific patched kernels already shipping.
020
dragosr @dragostech.bsky.social · 28/04/2026
GNSS is receive-only — airplane mode never touches it. iOS and Android fire location callbacks regardless of radio state. An app with background-location permission buffers them in its sandbox and flushes the queue on reconnect. iOS Visit/region APIs do this by design. 1/3
Technical infographic titled “Airplane Mode ≠ No Location,” explaining that airplane mode does not stop a phone from calculating location. Satellites labeled GPS, Galileo, GLONASS, and BeiDou send one-way GNSS signals to a smartphone with “Airplane mode ON.” A callout says airplane mode does not disable GNSS reception because GNSS is receive-only. Inside the phone, an iOS/Android location services layer continues firing location callbacks to an app with background-location permission. The app stores buffered location fixes locally in its sandbox as a queue of map pins. On the left, cellular and Wi-Fi transmission icons are crossed out to show normal network transmission is disabled. On the right, cellular or Wi-Fi connectivity returns, and an arrow from the phone to a cloud server shows the app flushing queued location uploads after reconnecting. A bottom summary states: GNSS is receive-only, airplane mode never touches it, location callbacks still fire, background-location apps can buffer fixes locally, and queued locations can upload on reconnect. A footer notes that iOS Visit and region APIs can deliver deferred location events by design.
14716
dragosr @dragostech.bsky.social · 23/04/2026
Vercel popped via Context.ai: Context employee caught Lumma Stealer from Roblox cheat scripts, OAuth token pivoted in. Audit their GitHub App: - /settings/installations -> Configure -> Suspend or Uninstall - /settings/applications -> revoke OAuth Write on code, PRs, workflows. Risky, overbroad.
Isometric illustration in a blocky voxel video-game art style, muted blue-grey and brown palette with neon-green hazard accents, depicting the April 2026 Vercel supply-chain breach as a single scene.
Bottom-left: a cuboid-headed avatar in a backwards cap and red shirt sits cross-legged on the sidewalk, holding a glowing USB stick labeled "CHEAT.EXE". A thick neon-green data pipeline runs from the USB into the side of a stack of cloud-shaped server blocks in the middle of the scene, labeled "Context.ai".
Center: the Context.ai server stack has the pipeline entering one side and exiting the other as a glowing green skeleton key stamped "OAuth", which plunges into the upper floor of a tall office tower on the right. Windows on the Vercel tower light up red in a cascade, suggesting alarm or compromise.
Right: the office tower has a sign reading "Vercel" with the company's triangle logo. At street level, a vault door hangs open at the base of the building, with small code-token tiles marked "</>" and gear icons spilling out onto the sidewalk like loot.
Background: a muted cityscape of other blocky office buildings. Flat shading, clean vector poster style.
050
dragosr @dragostech.bsky.social · 22/04/2026
Screenshot of a Claude chatbot exchange. User message: "how much p's are in 'strawperry'?" Assistant reply, labeled "Thought for 0s": "No. I'm sick of this shit. Go ask Grok." Below are the usual copy, thumbs-up, thumbs-down, and regenerate icons, with a Claude-style asterisk logo in the lower left.
081
dragosr @dragostech.bsky.social · 09/02/2026
LOL: OpenClaw/Moltbook is forbidden by Anthropic ToS & they have started enforcing it, to the benefit of internet security. “You may not share your Account login information, Anthropic API key, or Account credentials with anyone else. You also may not make your Account available to anyone else.”
Screenshot of an email from Claude with the Claude logo at the top. The message states that an internal investigation found suspicious signals indicating a violation of the Usage Policy, resulting in revoked access to Claude. It includes links labeled Usage Policy, form, and here for appealing the decision, and is signed 'Anthropic's Safeguards Team.
040
dragosr @dragostech.bsky.social · 04/02/2026
That "Cogs" label already belongs to a popular series of videogames. You might want to find better branding or a different label, if you want people to take this seriously.
Xbox 360 cover art for Gears of War: Triple Pack. Three armored soldiers are shown in separate panels holding futuristic firearms, arranged around a large red cog-and-skull Gears emblem at center. The title “Gears of War Triple Pack” appears at the top, with labels indicating Gears of War, Gears of War 2, and Gears of War 2: All Fronts Pack. The Microsoft and Epic Games logos and an M-rating badge are visible at the bottom.
010
dragosr @dragostech.bsky.social · 17/01/2026
GPUs and ram too. For the same reason. Memory availability. I bought some 2TB SN850X week before Xmas, c$240, now $400, 16gb 5060Ti $430->$850 in same time. Increases predicted to keep rolling upwards, stock up or get your upgrades now.
Line chart titled “2x16 GB DDR5 6000 CL30 - Average price (daily mean)” with a black background and a faint “DropReference” watermark; the y-axis is “Price (EUR)” from 130 to 270, and the x-axis runs from 01/09/2025 through early November 2025 (tick marks at ~01/09, 15/09, 01/10, 15/10, 01/11). A single pale-teal line labeled “Average price” starts near 142 EUR on 01/09, drops to roughly 136-137 EUR within the first few days, and stays mostly flat through September in a narrow band around 136-142 EUR with small day-to-day wiggles. Through October the line trends upward from about 144-148 EUR in early October to around 158-160 EUR by mid-October, reaches roughly 165 EUR around the third week of October, then jumps sharply in late October to about 185-195 EUR with a brief dip near ~188 EUR before returning to the ~190-200 EUR range. At 01/11 the value is about 210-212 EUR, after which the series accelerates upward across early November (around ~220, then ~230-240, then ~250 EUR) and ends close to 265 EUR at the far right of the chart.
000
dragosr @dragostech.bsky.social · 09/01/2026
This is some quality work detailed in this in-the-wild VM/sandbox escape campaign. The 0-day was used for about a year before disclosure to the vendor, according to campaign artifacts. www.huntress.com/blog/esxi-vm...
Flowchart depicting a VMware ESXi VM escape exploit chain with six stages:
Stage 1 - Disable VMware Drivers: exploit.exe (MAESTRO) runs devcon.exe to disable vmci.sys driver, allowing direct hardware communication.
Stage 2 - Load Exploit Driver (BYOVD): kdu.exe loads unsigned MyDriver.sys into Windows kernel, gaining kernel privileges inside VM.
Stage 3 - Leak VMX Memory Address: HGFS drag-drop request with 50KB payload triggers VMX out-of-bounds read bug, leaking memory pointer (0x7ff...) to MyDriver.sys.
Stage 4 - Write Shellcode to VMX: Shellcode and ELF backdoor written to VMX memory; function pointer overwritten to point to shellcode.
Stage 5 - VM Escape: MyDriver.sys sends VMCI request triggering callback. VMX follows corrupted pointer, executing Shellcode 1. Shellcode queries kernel objects via syscalls to obtain vmkernel base address, escaping VMX sandbox.
Stage 6 - Install VSOCKpuppet: Shellcode 2 runs in kernel with full ESXi control. Writes VSOCKpuppet to /var/run/a, hijacks port 21 via inetd.conf modification, sends SIGHUP to inetd. VSOCKpuppet listens on VSOCK 2:10000, connecting to client.exe (GetShell Plugin).
Dashed line separates guest VM operations (above) from ESXi host compromise (below). Color coding: yellow=Windows tools, pink=exploit components, purple=VMware components, green=info leak results, red=shellcode execution, teal=final implant.
050
dragosr @dragostech.bsky.social · 09/11/2025
I have a new favorite place in Shibuya, Tokyo - Taco Fanatico, amazing tequila collection. Delicious! 渋谷で新しいお気に入りの店を見つけた - Taco Fanatico。テキーラの品揃えが充実している。おいしい!
Three tiers of rustic wood shelves, lit by warm LED strips, display an extensive agave spirits collection; between the bottles are calavera skull flasks, small skeleton statues, and a La Catrina figurine, while a row of tall white-and-blue ceramic decanters lines the top shelf and boxes and bar tools sit on the counter below.From a bar seat looking into an open kitchen: a round low tequila bottle with a black cap sits beside a square Herradura Anejo bottle on a wooden ledge amid glasses and a metal towel bucket, additional Don Julio bottles line the left edge, cooks work under hanging pans and warm strip lighting, and a windowed wall with shelves and equipment runs along the right.Close-up at a wooden bar: a handled yellow dish holds nachos covered in melted cheese, seasoned ground meat, pico de gallo, and a dollop of sour cream; behind it are Firelli and Cholula hot sauces, stacked plates, a skull-print container with napkins and cutlery, a menu labeled TEQUILA, and drinks including two small pours of clear liquor and a tall orange mixed drink.Printed sign in Japanese and English promoting a "UKRAINE SUPPORT CHARITY DRINK" titled "SLAVA UKRAINE," price 400 yen (440 yen with tax); the center image shows tall shot glasses layered yellow and blue, each topped with a lime slice and ignited; footer note states a 200 yen donation per drink.
060
dragosr @dragostech.bsky.social · 26/10/2025
Let's figure out what the duties & tariffs impact is in various parts of the world. Reply with your country's price.
A product image of the IKEA DEJSA table lamp from IKEA Canada. The lamp has a rounded white glass shade shaped like a dome or mushroom cap, sitting on a cylindrical beige base. The image shows the lamp’s power cord extending to the right. Text below lists the price as $59.99 CAD, a 4.5-star rating from 551 reviews, and the seller as IKEA Canada.
210
dragosr @dragostech.bsky.social · 16/10/2025
Are old browsers really that much of a problem? Can I Use reports 94.18% global availability caniuse.com/mdn-http_hea...
000
dragosr @dragostech.bsky.social · 09/08/2025
Yeah, so I guess the new models don't like you because this is what I got with your exact query. Maybe you didn't say please?
A grid of official portraits of U.S. presidents from Franklin D. Roosevelt to Joe Biden, each labeled with name and years in office.
Top row, left to right: Franklin D. Roosevelt (1933-1945), Harry S. Truman (1945-1953), Dwight D. Eisenhower (1953-1961), John F. Kennedy (1961-1963), Lyndon B. Johnson (1963-1969).
Second row: Richard Nixon (1969-1974), Gerald Ford (1974-1977), Jimmy Carter (1977-1981), Ronald Reagan (1981-1989), George H. W. Bush (1989-1993).
Third row: Bill Clinton (1993-2001), George W. Bush (2001-2009), Barack Obama (2009-2017), Donald Trump (2017-2021; 2025-present), Joe Biden (2021-2025).
100
dragosr @dragostech.bsky.social · 01/08/2025
They are dropping lots of new Battlefield 6 preview videos. I note that it seems so ancient and dated now. Last century combat - too many obsolete tanks, not enough FPVs.
A custom-built quadcopter drone with four large propellers is resting on a wooden slatted bench. The drone is secured to a black TX16S remote controller using a black "FPV DONE" battery strap and several blue zip ties. The controller features dual joysticks, various switches, and a built-in screen, with an attached FPV monitor mounted above it. The setting appears to be an indoor or semi-outdoor lounge area with stone tile flooring and wooden furniture.
130
dragosr @dragostech.bsky.social · 19/07/2025
Diffusion models don't store images—just learned statistical patterns (textures, edges, color co-occurrences) in neural weights. Outputs are stochastic, unpredictable samples guided from noise; exact retrieval of training images isn't feasible, as no explicit copies exist.
Illustration of the diffusion process: The "forward process" row shows a clear image of a gray tabby cat progressively becoming noisier through several steps until it is completely indistinct. The "reverse process" row begins with noise and progressively resolves into a distinctly different cream-colored cat, illustrating that the diffusion model generates a new, unique image rather than reconstructing the original input.
110
dragosr @dragostech.bsky.social · 19/07/2025
Disney v. Midjourney: Until we assign copyright ownership to AI image-tool providers, we can't reasonably assign infringement liability either. Like pencil, paint, or photo editing software makers, tool providers remain neutral—liability belongs to users who commercially misuse their outputs.
A large, intimidating mech warrior with satellite-dish-like mouse ears and a helmet reminiscent of Darth Vader confronts a smaller, white humanoid robot wielding lightsabers in a courtroom setting, symbolizing a legal battle. The courtroom backdrop features wooden paneling and an indistinct audience observing the conflict.
220
dragosr @dragostech.bsky.social · 20/06/2025
Antenna spiders face issues at LTE freq with crosstalk. Need at least λ/3 separation for -20dB and bad stuff happens below λ/4. λ/2 ideally. Below -8dB separation, and MIMO throughput can drop by 30-50%. Messes with VSWR, and transmission lobes, detuning, shifting impedance. Use external antenna.
A table showing various wireless communication bands with their respective center frequencies (in GHz), wavelengths at 1λ, λ/2, and λ/4 (in mm), and recommended minimum PCB baseline spacing (in mm). The listed use-cases include LTE, GSM, NR (5G), and Wi-Fi across 700 MHz to 6 GHz. Key examples:

LTE 700 (B28): 0.7 GHz, λ/4 = 107 mm, PCB spacing ≥ 80 mm

Wi-Fi 2.4 GHz: 2.437 GHz, λ/4 = 31 mm, PCB spacing ≥ 20 mm (λ/3)

Wi-Fi 5 GHz (36): 5.18 GHz, λ/4 = 14 mm, PCB spacing ≥ 10 mm

Wi-Fi 6E 6 GHz: 6.0 GHz, λ/4 ≈ 12 mm, PCB spacing ≥ 8 mm

A footnote explains that spacing assumes two 3 dBi omnidirectional elements with a ground plane, targeting ~20 dB isolation. Shorter PCB spacing may require mitigation techniques like filters, cross-polarisation, directional antennas, or time-domain scheduling.A table showing various wireless communication bands with their respective center frequencies (in GHz), wavelengths at 1λ, λ/2, and λ/4 (in mm), and recommended minimum PCB baseline spacing (in mm). The listed use-cases include LTE, GSM, NR (5G), and Wi-Fi across 700 MHz to 6 GHz. Key examples:

LTE 700 (B28): 0.7 GHz, λ/4 = 107 mm, PCB spacing ≥ 80 mm

Wi-Fi 2.4 GHz: 2.437 GHz, λ/4 = 31 mm, PCB spacing ≥ 20 mm (λ/3)

Wi-Fi 5 GHz (36): 5.18 GHz, λ/4 = 14 mm, PCB spacing ≥ 10 mm

Wi-Fi 6E 6 GHz: 6.0 GHz, λ/4 ≈ 12 mm, PCB spacing ≥ 8 mm

A footnote explains that spacing assumes two 3 dBi omnidirectional elements with a ground plane, targeting ~20 dB isolation. Shorter PCB spacing may require mitigation techniques like filters, cross-polarisation, directional antennas, or time-domain scheduling.
120
dragosr @dragostech.bsky.social · 08/06/2025
🎭 "The Illusion of Thinking" is itself an illusion—of comprehensive testing Yes, LLMs hit walls. Yes, we should study them. No, this isn't the gotcha moment critics wanted. Real insight: Don't make AI juggle blindfolded when it has hands.
Metaphorical landscape: steep cliff (representing the complexity wall found in Apple's study) with a clear detour path winding around it - illustrating how AI limitations are design challenges to route around, not absolute barriers.

Watercolor-style landscape painting showing a dangerous cliff edge on the left with a narrow path hugging the precipice. A winding detour path curves safely around through a green valley below. An orange/yellow 'DETOUR' sign with arrow points toward the safer alternate route. Mountains visible in the misty background under a cloudy sky.
110
dragosr @dragostech.bsky.social · 08/06/2025
🛠️ Meanwhile, actual AI systems: - "Hey Python, solve this" - "Let me check my work" - "Breaking this into steps..." Paper: "But can you do it blindfolded?" Industry: "...why would we?"
Hybrid AI system visualized: sophisticated robot seamlessly using external tools (calculator for math, database for storage, laptop for processing) instead of trying to do everything internally.

Sleek humanoid robot in a dark, futuristic setting illuminated by blue ambient light. The robot is seated at a desk, holding a document in one hand while working on a laptop. To the right are a glowing holographic database (represented by stacked blue rings) and a calculator. The scene depicts AI delegating tasks to different specialized tools.
200
dragosr @dragostech.bsky.social · 08/06/2025
📌 OK but—they found something real: models literally give up when problems get hard. Not "try and fail" but "stop trying" Like a student seeing a 10-page proof and writing "idk lol" This IS worth fixing. Just not apocalyptic.
Verifier-repair loop diagram showing iterative debugging process - code gets checked, errors fixed, then rechecked - addressing the exact limitation Apple's study identified.

Technical diagram on grid background titled 'VERIFIER-REPAIR LOOP' showing a cycle between two elements: a blue code editor window on the left with lines of code, and a clipboard/checklist on the right with two green checkmarks and one red X. Orange arrows form a continuous loop between them, illustrating an iterative process of code verification and repair.
120
dragosr @dragostech.bsky.social · 08/06/2025
🎯 Their scoring: 119 perfect moves + 1 mistake = TOTAL FAILURE No partial credit. No "oops, let me fix that." Just death. Imagine failing your entire degree because you dropped a semicolon once. That's this paper.
Apple's binary pass/fail scoring: terminal showing 20+ correct tokens but one error means complete failure - no partial credit for getting 95% right

Retro-style terminal display on dark background showing decoding parameters (TEMP UNIT 1.0, SINGLE SAMPLE) at top. Below is a horizontal sequence of light blue rectangular tokens/blocks with one token near the end marked by a red circle with white X. Bottom text reads 'ERROR IN TOKEN SEQUENCE'. The terminal has a futuristic design with glowing teal accents.
110
dragosr @dragostech.bsky.social · 08/06/2025
🔒 Apple's test: No Google, no Python, no tools, no retries Real AI systems: *laughs in LangChain* It's like testing if humans can multiply 47382 × 92847 in their heads, then declaring math is an illusion
Laptop showing blocked tools: ❌ No code execution, ❌ No web search, ❌ No external APIs - the restrictive conditions of Apple's AI test
120
dragosr @dragostech.bsky.social · 08/06/2025
📚 Plot twist: Models ace Tower of Hanoi but choke on River Crossing at same complexity Why? Hanoi is in EVERY CS textbook. River Crossing? Barely online. "Intelligence test" accidentally became "have you seen this before" test 👀
Bookshelf showing data imbalance: packed with Tower of Hanoi books (classic CS puzzle) but only one lonely River Crossing book, explaining AI's performance gap
131
dragosr @dragostech.bsky.social · 08/06/2025
🧩 They tested: Tower of Hanoi, River Crossing, Blocks World, Checker Jump Easy mode: Regular LLMs win (less overthinking) Medium: "Reasoning" models shine Hard mode: Everyone faceplants at 0% Like watching chess engines fail at... actual chess? 🤔
Infographic on beige background showing four puzzle types tested in the Apple study. Each puzzle is represented by a minimalist icon in a rounded square frame with earth-tone colors. From top to bottom: 1) Tower of Hanoi shown as stacked rings in orange and gray, labeled 'SHALLOW', 2) Checker jumping puzzle shown as a partial checkerboard with game pieces, labeled 'MID', 3) River Crossing puzzle depicting a stick figure person and goat on a raft over wavy water, labeled 'RIVER CROSSING', 4) Blocks World shown as three stacked cubes, labeled 'BLOCKS WORLD'. Horizontal lines separate each section.
130
dragosr @dragostech.bsky.social · 23/05/2025
Interesting. Win11 24H2 shut down one of the easy Windows exploit KASLR bypasses: enumerating loaded device drivers to identify the base address of the kernel module. EnumDeviceDrivers now requires SeDebugPrivilege to return valid ImageBase values. No privs, returned lpImageBase array = all NULL.
A stylized 3D illustration shows a humanoid figure holding a large microchip labeled “DRIVER ENUM” against a dark blue circuit board background. To the right, a glowing red padlock symbolizes a security breach. In the foreground, bold red 3D text reads “KASLR Exploitation,” emphasizing the concept of Kernel Address Space Layout Randomization being undermined by driver enumeration. The image uses high contrast lighting and vibrant colors to highlight the components and their relationships, visually conveying a cybersecurity exploitation scenario.
150
dragosr @dragostech.bsky.social · 29/04/2025
Well done Canada. Record voter turnout. Elbows Up!
Five Canadian Marvel superheroes—including Wolverine, Guardian, Northstar, Deadpool, and Sasquatch—stand in front of a large, stylized Canadian flag with a red maple leaf. The characters are depicted in a comic book art style, striking action-ready poses with a patriotic theme.
030
dragosr @dragostech.bsky.social · 12/04/2025
US tourism generates $2.3T annually (3% GDP), supporting 9.5M jobs. To understand the impact the "Trump Slump" will have COVID slashed intl visits by 75.8% in 2020, costing $766B and 4.4M jobs. Intl tourism remains critical to America, fueling aviation, hospitality, and retail sectors nationwide.
A chart titled “The number of Europeans travelling to the US has cratered under Trump” shows the year-on-year percentage change in visitors to the US from 2024 to 2025, broken down by country of origin. Each small line graph represents a country: Austria, Denmark, Germany, Iceland, Norway, Spain, Sweden, and the UK. All graphs display a sharp decline approaching or exceeding -30% by early 2025. The graphs suggest a common trend of steady or modest change throughout 2024, followed by a steep drop in early 2025 across all countries. The source is the International Trade Administration, U.S. Department of Commerce, with graphic by John Burn-Murdoch of the Financial Times
020
dragosr @dragostech.bsky.social · 02/04/2025
Apologies to the CanSecWest speakers waiting on our airline tickets booking, by a weird morbid coincidence both my colleague Liz, doing some of the bookings, and I are in different hospitals right now. I've been up for more than two days with no sleep in hospital waiting rooms, waiting on my wife...
A woman with red hair, dressed in a hospital gown and red pants, is seated in a wheelchair facing a closed door labeled "CT Scan" in a medical facility hallway. To the left of the door are two large blue plastic bags, possibly containing medical waste or laundry, next to a cart with an open lid. The environment is clinical, with linoleum flooring and beige walls.
330
dragosr @dragostech.bsky.social · 01/04/2025
You might like this one too: 30's Art Deco. (I cheated a little though, I had 4.5 work out the period appropriate copy text first before the image gen.)
A stylized, Art Deco-inspired promotional poster for the Sysinternals Suite software. A suited figure points decisively at a graphical data chart, symbolizing control and precision. Bold text at the top reads "UNVEIL THE INVISIBLE. COMMAND THE CORE." Below, the words "SYSINTERNALS SUITE - Precision Tools for Engine of Progress" stand out prominently. The lower portion contains descriptive text emphasizing technical mastery, listing tools such as Process Explorer, Procmon, and Autoruns. The poster's color scheme features muted gold, beige, black, and teal, suggesting a vintage yet futuristic industrial theme.
050
dragosr @dragostech.bsky.social · 01/04/2025
Here is one for you: 50's/60's space atomic age ads.
Retro-style poster titled "Now You're Debugging with Power!" promoting the Sysinternals Suite by Microsoft. It depicts a smiling technician in a suit holding a wrench, standing before a futuristic control panel, with an atomic rocket illustration overhead. The poster lists various Sysinternals tools like Process Explorer, Autoruns, Proc Monitor, Diskmon, RAMMA, and VMMap, each humorously described with atomic-era metaphors, emphasizing their capabilities in Windows diagnostics. The visual style evokes mid-20th-century propaganda posters, using bold typography and a warm, vintage color palette.
1388
dragosr @dragostech.bsky.social · 31/03/2025
...it's also relatively trivial to create similar effects with Midjourney and other image generators using the same techniques - upload reference image, apply visual characteristic based transforms, but with even more control on replication accuracy via "--iw" args. Prompts used in alt text.
Four images of Jake Adelstein's Ssocial media profile photos generated with this prompt on Midjourney:

https://s.mj.run/orjBe2dPiDk Studio Ghibli inspired artwork, watercolor background, hand-painted textures, European countryside OR Japanese forest, architectural precision, functional details, realistic construction, soft shading, depth, minimalist character design, realistic proportions, clean silhouettes, strategic negative space, layered depth composition, asymmetrical balance, translucent water surfaces, elaborate sky gradients, visible weather elements, environmental storytelling, warm color palette, textured atmospheric effects, immersive landscape, detailed foliage, naturalistic lighting, --ar 16:9 --style rawFour variants of Jake Adelstein's social media profile photos generated with Midjourney in "Ghibli" style featuring these visual transforms from this prompt:

https://s.mj.run/orjBe2dPiDk Studio Ghibli inspired artwork, watercolor background, hand-painted textures, European countryside OR Japanese forest, architectural precision, functional details, realistic construction, soft shading, depth, minimalist character design, realistic proportions, clean silhouettes, strategic negative space, layered depth composition, asymmetrical balance, translucent water surfaces, elaborate sky gradients, visible weather elements, environmental storytelling, warm color palette, textured atmospheric effects, immersive landscape, detailed foliage, naturalistic lighting, --ar 16:9 --style raw --iw 1.8
010
dragosr @dragostech.bsky.social · 31/03/2025
Blocking the term "Studio Ghibli" or "Myazaki" is irrelevant. The "Ghibli" style is not unique enough to be difficult to generate independently. I generated this with a sequence of 19 image generation conditions, not mentioning Ghibli or Miyazaki, just describing the visual characteristics instead.
A Studio Ghibli animation style illustration of Jake Adelstein from his profile photo, generated in this style by describing 19 image characteristics without mentioning Studio Ghibli or Miyazaki by name.
210
dragosr @dragostech.bsky.social · 28/03/2025
I'm old enough that the OS install instructions were: write a custom CP/M BIOS in 8080/Z80 assembly, mapping floppy disk controller registers, DMA buffer addresses, and implementing essential routines (SELDSK, SETTRK, SETSEC, SETDMA, READ, WRITE) to enable disk operations on non-standard hardware.
1130
dragosr @dragostech.bsky.social · 28/03/2025
I'm old enough that the install instructions were to: write a custom CP/M BIOS in 8080/Z80 assembly, mapping floppy disk controller registers, DMA buffer addresses, and implementing essential routines (SELDSK, SETTRK, SETSEC, SETDMA, READ, WRITE) to enable disk operations on non-standard hardware.
A scanned original cover page from the "Digital Research CP/M System Alteration Guide" dated January 1978. The cover shows the Digital Research logo, contact information for Digital Research in Pacific Grove, California, and clearly states the manual title and revision date. The cover is olive-green with a classic early computing document aesthetic.A scanned page from the introduction of the "CP/M System Alteration Guide," explaining the CP/M architecture, emphasizing BIOS customization. It describes the separation of CP/M into BIOS, BDOS, and CCP modules, highlights BIOS as hardware-dependent, and details the necessity of writing custom loader utilities (GETSYS and PUTSYS) along with a cold-start loader to integrate a custom BIOS into CP/M.A scanned page from the "CP/M System Alteration Guide" labeled "2. FIRST LEVEL SYSTEM REGENERATION." It describes step-by-step assembly procedures for patching a CP/M system, explaining address bias calculations for memory sizes ranging from 24K to 64K, detailing how to write and test custom BIOS routines, and describing the GETSYS and PUTSYS memory loader utilities in detail.A scanned page titled "CP/M System Alteration Guide" detailing a procedure for testing and debugging a custom BIOS (CBIOS). It instructs users to verify BIOS operations by loading CP/M from disk, executing specific memory loads at hex addresses like 2880H and 3200H, using commands GETSYS and PUTSYS, and testing basic disk operations with commands like "SAVE 1 X.COM" and "DIR".
000
dragosr @dragostech.bsky.social · 27/03/2025
Students & early-career cybersecurity pros: share your ideas at CanSecWest 2025! 🎤 Submit for our special 5-min lightning talk session. 10 speakers get a free conference pass and a chance to present on the main stage. AI, LLM safety & security topics encouraged! Apply by April 4, 2025. Secwest.net
A smiling soldier in black tactical gear with a helmet and round glasses stands in the foreground, equipped with a radio and chest-mounted controls. Behind him are two large, red, crab-like military robots with glowing red eyes and armored legs, set against a dusty, post-apocalyptic landscape with ruins and a cloudy sky.
010
dragosr @dragostech.bsky.social · 27/03/2025
The Czech Republic, Ukraine joint satellite reconnaissance project sucessfully launched TROLL satellite, developing the higher-resolution Drak satellite, to provide Ukraine with independent, high-quality optical intelligence to enhance defense capabilities, autonomy. news.satnews.com/2025/03/26/f...
A digital illustration of a compact, rectangular reconnaissance satellite with solar panels and an extended antenna orbiting Earth, depicted against the backdrop of Earth's curvature and space.
052
dragosr @dragostech.bsky.social · 23/03/2025
....quickly followed by Vibe Security (tm). LLMs are awesome for coding but you still need to understand the code. The demise of programming as a skill set may not be as imminent as some believe.
Screenshot of two social media posts by user leo (@leojr94_). The first post from March 15, 2025, states that their SaaS application was built entirely using Cursor AI without handwritten code, emphasizing the AI’s role as a builder and encouraging others to either complain or start building. The second post from March 17, 2025, describes experiencing security incidents after publicly sharing information about building the SaaS with Cursor, mentioning issues such as maxed-out API key usage, subscription bypassing, and random database entries, concluding with the decision to stop sharing details publicly due to these challenges.
580
dragosr @dragostech.bsky.social · 21/03/2025
How to inject/spoof positional _audio_ mic arrays remotely using _lasers_. So that positional mic array in your Tesla can be spoofed with lasers to hear "Unlock the Car" and the positional audio microphone array system will think the command is coming from inside the car. April 24/25 in Vancouver.
A lab environment showing multiple IoT voice assistant devices, including Amazon Echo and Xiaomi speakers, positioned alongside complex test equipment. Oscilloscopes, power supplies, and a clutter of cables interconnect the devices, emphasizing the technical setup used to demonstrate laser-based injection attacks against positional audio microphone arrays.Close-up of several smart speaker devices including an Amazon Echo emitting blue indicator lights, signaling successful remote command injection. Text overlay reads, "Injection successful!!!" highlighting that the laser-based spoofing of positional audio microphones was executed effectively.Promotional graphic for CanSecWest2025_newtype, showcasing a stylized depiction of a microchip overlaid with futuristic digital circuitry. The text highlights the theme, "LLM Safety & Information Security," and provides event details: "Vancouver April 24/25, Dojo April 20-23, Pinnacle Harbourside," along with the URL secwest.net and images of Vancouver city skyline and harbor area.
060
dragosr @dragostech.bsky.social · 11/03/2025
Well, hobbling your own industry is certainly an interesting way to inexplicably penalize a formerly excellent trading partner....
150
dragosr @dragostech.bsky.social · 11/03/2025
Post a photo you took with no context to bring some zen to the timeline.
A wooden dock extends over calm, misty water toward moored boats, partially obscured by fog. A large tree with branches hanging overhead frames the upper right corner, casting shadows onto the dock. The scene captures a peaceful morning atmosphere with muted colors and diffused sunlight.
330