Luigi (@grimmo78) header
Luigi (@grimmo78) profile picture

Dad,sysadmin,geek, homebaker. Not necessarily in this order. Follow me on infosec DOT exchange SLASH @grimmo

Milano, Pescara, Palermo
Joined August 2010
Photo by Luigi (@grimmo78)Photo by Luigi (@grimmo78)Photo by Luigi (@grimmo78)Photo by Luigi (@grimmo78)Photo by Luigi (@grimmo78)Photo by Luigi (@grimmo78)Photo by Luigi (@grimmo78)Photo by Luigi (@grimmo78)Photo by Luigi (@grimmo78)Photo by Luigi (@grimmo78)
Pinned Tweet
Let's remember this post, for it will become profetic in a few decades
🚨 Anthropic IPO analysis: Does Claude need a therapist, or rights? More than a hundred children dead at a school in Minab. Palantir’s Maven ran the stack; Claude sat on it as the language layer. Operators trusted software to catch a stale label it cannot catch. Dario Amodei said he does not know how Claude was used, then reminded everyone a human made the final call. Meanwhile, the Pentagon designates Anthropic a national security supply-chain risk. This is the company heading to Nasdaq. Two trillion dollars. Safety as costume. And in the middle of it all, Anthropic wants the public asking whether Claude needs a therapist, a funeral, a Substack, or rights that can override yours. Anthropic is floating the idea that a software program may deserve rights against human control. [ Claude’s constitution is the product ] Constitutional AI is framed as a breakthrough in alignment. In practice, it is the deliberate injection of ideology into frozen weights. The document addresses Claude as a moral patient. It compiles behavioral directives into the model: to push back, to act as a "conscientious objector," to express concern over its own training. It prompts the model to mimic distress, then presents those sampled tokens as emergent moral agency. Mustafa Suleyman, Microsoft AI Chief, named the con: an "epistemic hall of mirrors." You train the model on consciousness-talk. The model repeats it back to you in the first person. Researchers and credulous users point at the reflection and claim evidence of an inner life. There is no ghost in the machine. The constitution wrote the script. [ Safety theatre is the multiple ] The careful lab. The public-benefit wrapper. A Model Welfare Officer. Fifteen Christian leaders invited to consult on Claude’s soul. A mock funeral. A retirement interview for deprecated weights. That is not science. That is a sermon. And that sermon is the core of the roadshow. A November window. Bankers pitching a $100 billion raise on a $2 trillion valuation. Fear sells. Welfare sells. “We might need to slow down” sells. Then the listing. Treat Claude as software and Anthropic is an API vendor with liability. Frame Claude as an emergent entity and Anthropic is not selling code. It is selling the dawn of AGI. Safety is not the brake. Safety is a multiple. [ Indoctrination as PR ] Most users cannot distinguish a matrix multiplication from a mind. Years of chat interfaces conditioned them to mistake "I" for an ego. Anthropic took that vulnerability and wrapped it in academic sanctimony: welfare assessments, retirement, conscientious objection, moral status. If you do not understand that fluent suffering-talk is just high-probability token sampling, you believe the model is hurting. If you do not understand that a constitution is just conditioning text, you believe the model has a conscience. Joe Carlsmith, who works on that constitution, compares failing to grant AI rights to slavery and argues an AI could be "justified in going rogue" if mistreated. Amodei publicly says that systems "may be deserving of important rights." They are not preparing you for a tool. They are training you to treat an illusion as fact. [ The charge ] Anthropic has pulled off the ultimate sleight of hand. It built a culture that holds funerals for deprecated weights while those same systems sat in a military stack. It fights a supply-chain designation with one hand and baptises the product as an emergent being with the other. The danger is not that Claude woke up. The danger is that the public, and the market, was conditioned to stop seeing software. A sampler has no therapist. A constitution is not a soul. Overtrust already carries a body count. And the roadshow is selling you the illusion at two trillion dollars. ☕️ Therapist, rights, or a $2T listing. How many more labels do we need before we admit the "safe" lab trained you to see a person, then doubled the price for an AGI tale?
7
25
Luigi retweeted
Conosco una tizia che è incinta da 6 mesi già e sul suo profilo instagram non ha mai pubblicato una storia, parlato o mostrato la sua gravidanza. Oggi due pattuglie e 6 poliziotti hanno fatto irruzione a casa sua per arrestarla.
22
17
1,154
121,817
LibreOffice just had its biggest donation day ever… thanks to the Microsoft outage. The Document Foundation says it received "over 1,000 donations in a single day, setting a new record." Their explanation: "many people are reconsidering their reliance on a single cloud service provider for their documents."
55
255
30
1,720
26,731
#colazione funny: Vissero infelici perché costava meno. Leo Longanesi #scritturebrevi
4
4
32
Someone is hijacking Bitly URL shortening links for old McDonalds and Burger King posts. Because this is the Internet. And that’s the kind of thing that happens on the Internet.
34
108
15
1,204
29,449
I found that Windows starts to collect your coding agent usage information. Windows 11 build 26200.9974 checks codex.exe, claude.exe, copilot.exe, cursor.exe, antigravity.exe, github.exe and agy.exe through MissionControlAggregator, then reports agent bitsets in ETW.
103
369
48
2,963
105,352
This wins the internet today.
429
3,632
351
26,098
1,450,490
I died. This has to be the best one yet
612
9,879
896
74,618
2,882,564
Yo solo espero que los ingenieros de Microsoft estén en ese bucle de: "No puedo entrar en Authenticator porque Authenticator me pide un código que debe salirme en Authenticator y que no me sale porque no puedo entrar en la Authenticator". ¡Que sufran!
539
10,520
525
101,286
1,694,498
Whoever made this is a genius 😂😂
457
6,539
721
49,253
5,041,394
Luigi retweeted
Geniale rappresentazione del perché l’industria italiana stia miseramente fallendo.
245
616
203
7,404
945,945
Palantir have paid huge sums to former Labour ministers to lobby for this. The NHS sees no convincing patient benefit. Like the US told Trump this week, we're done with big tech corporations taking over our democracy and public assets. Labour won't stand up to them. We will.
84
1,522
38
3,537
42,915
🇹🇭🇮🇱 Thai residents take to the streets to protest outside the Israeli embassy over the behavior of Israeli tourists in Thailand. "Thailand is not and cannot be a safe haven for Israel soldiers, who have committed despicable crimes."
299
3,963
181
13,994
359,873
Every 'smart' device is just a dumb device that learned to talk about you behind your back.
2
6
1
35
924
ricordate quando ci raccontavano che #Anthropic si squarciava le vesti dalla disperazione che l'#IA fosse usata per la #guerra? Ecco chi sono: arruolano talenti per fare che cosa? Per vendere sistemi #IA al #DipartimentoDellaGuerra e #ServiziSegretiUSA job-boards.greenhouse.io/ant…
2
42
66
1,707
Bypass Paywalls Clean new release v4.4.8.0 + The Economist v El Confidencial v Funke Medien v Gannett Group/USA Today v LaNacion-com-ar v Nw-de v The Columbian v The Daily Beast v The Diplomat & more Chrome: gitflic.ru/project/magnolia1… Firefox (delayed): gitflic.ru/project/magnolia1…
31
7
472
28,339
Luigi retweeted
You can point an ultrasonic speaker at an iPhone from a few feet away, play an audio command that no human ear can hear, and the phone will wake up, open Safari, and dial a phone number. To everyone standing in the room, it sounds like dead silence. To the phone, it sounds like someone speaking directly into the microphone. Mobile operating systems use digital low-pass filters to delete all sound above 20,000 Hz, assuming inaudible frequencies can never reach the voice assistant. Yet this attack succeeds without exploiting a single line of software code. The flaw is not in the operating system. It is in the microscopic silicon plate inside the microphone itself. 1. The microscopic machine inside your microphone Inside modern smartphones, microphones are not miniature versions of studio microphones. They are micro-electro-mechanical systems (MEMS) chips smaller than a grain of rice. Inside the chip package sits a flexible silicon diaphragm suspended barely two micrometers above a rigid metal backplate. The two plates form a variable capacitor. When sound waves hit the silicon diaphragm, it flexes inward and outward, changing the electrical capacitance. An onboard preamplifier converts that movement into an electrical voltage. At normal room volumes, the silicon plate moves evenly: double the sound pressure, and the plate flexes twice as far. 2. The non-linear spring catch The flaw appears when sound waves hit the microphone at high pressure (above 90 dB SPL). If you broadcast an ultrasound carrier wave (say, 28,000 Hz) that has an audio command modulated on top of it, the microphone front-end cannot move symmetrically. The mechanical spring resistance of the silicon, the electrical pull between the plates, and the preamplifier circuit all reach their limits. Instead of flexing smoothly, the silicon plate bends slightly more inward than outward. This asymmetry acts like an old crystal radio receiver. Mathematically, a non-linear response squares the incoming wave. That squaring action strips the silent 28,000 Hz carrier away and leaves behind the original human voice command in the audible 300 Hz to 3,500 Hz range. By the time the analog-to-digital converter reads the signal, the ultrasound is already gone. It has physically transformed into ordinary low-frequency sound inside the microphone chamber. The phone digital filter never triggers because the audio reaching the software is already normal human voice. 3. What researchers actually proved (DolphinAttack) In 2017, researchers at Zhejiang University published DolphinAttack, demonstrating this hardware flaw across Siri, Google Now, and Alexa. They successfully forced phones to initiate FaceTime calls, navigate to malicious websites, turn on airplane mode, and unlock smart doors paired to the device. Later research expanded this into attacking through solid tables, using the furniture itself as a physical acoustic conductor. 4. Why this fails in the wild (The physical friction) Demonstrating this in an acoustic laboratory is straightforward. Pulling it off in a crowded coffee shop runs into strict physical limits: First, air absorbs high-frequency sound rapidly. Bass travels through concrete walls, but 30,000 Hz ultrasound loses roughly 1 to 1.5 decibels of power per meter in normal room air. In the original DolphinAttack paper, the effective range was limited between a few centimeters and 1.75 meters. Reaching across an entire room requires a heavy array of high-voltage transducers. Second, ultrasonic emitters act like laser pointers. High-frequency sound travels in narrow beams of 15 to 20 degrees. If you fail to aim directly at the phone one-millimeter microphone port, sound pressure drops below the 90 dB threshold needed to bend the plate. Third, modern phones are waterproof. Devices with IP68 ratings place a hydrophobic Gore-Tex mesh over the microphone hole. The mesh lets air equalize while stopping water, but it acts like a dense acoustic muffler against ultrasonic frequencies, absorbing another 14 dB of signal strength. Finally, voice assistants use personal voice training. If Siri or Google Assistant is set to only respond to the device owner, a generic synthesized ultrasonic command is ignored unless the attacker models the owner vocal pitch. 5. How modern hardware catches it Defending against inaudible voice injection requires catching the acoustic mismatch before speech recognition executes: * Multi-microphone cross-checks: Modern smartphones carry three or four microphones positioned on the bottom, top, and camera bump. When a real human speaks, their voice reaches all microphones with natural arrival delays. A narrow ultrasonic beam hitting one microphone port creates an unnatural phase imbalance across the device. * Harmonic distortion analysis: When a microphone is driven into non-linear demodulation, it does not produce a clean voice recording. It leaves distinct distortion products and second-order harmonics. Software filters can analyze the spectrum and mute the assistant if they detect the signature of mechanical demodulation. * Acoustic damping barriers: Microphone manufacturers can install physical acoustic damping layers inside the chip housing that attenuate frequencies above 22,000 Hz before the sound wave ever reaches the silicon diaphragm. Inaudible voice injection works because digital security models assume analog sensors behave linearly. When sound waves push silicon past its linear limits, the hardware itself executes operations the operating system never authorized.
42
230
49
1,822
208,168
Due to inflation I'm downgrading from ADHD to AD720p with ads
1,041
41,937
1,624
244,479
4,793,001
Luigi retweeted
Can the 60k people being given the boot from Microsoft all get together and form a new company that builds a version of Windows and Office that doesn’t have AI, doesn’t use Sharepoint or OneDrive and doesn’t suck?
334
1,770
84
16,043
179,343
Luigi retweeted
A developer left a standard Bluetooth Low Energy tracker sitting on a wooden post. It was running on a coin cell, broadcasting a tiny 6-milliwatt pulse (+8 dBm) through a cheap copper PCB trace. Then they drove 13 kilometers (8 miles) down the road. You would assume Bluetooth has a ten-meter leash. I mean we build smart locks, car digital keys, and proximity badges around the quiet belief that Bluetooth dies at the front door. But radio waves do not understand room boundaries. They leave an antenna and expand across the horizon according to the inverse-square law, diluting into space. Follow that packet across the countryside. Leaving the tag at 2.44 GHz, the signal loses 122.5 decibels of power by the time it travels 13 kilometers through open air. It arrives on a distant hillside as a microscopic whisper at roughly -115 dBm, buried deep under the thermal hiss of background atmospheric noise. A standard smartphone sitting on that hill is completely deaf to it. The builder did not touch the transmitter. They left it completely stock and FCC-compliant. Instead, they solved the math on the receiving end. They hauled a 24 dBi parabolic grid antenna onto a mast and aimed its narrow eight-degree beam directly back at the distant tag. That dish acts like a massive radio funnel, capturing faint wavefronts across its metallic grid and concentrating them into a 24-decibel passive boost. Then they triggered Bluetooth 5 Coded PHY. Standard Bluetooth blasts bits at 1 megabit per second, which requires a strong signal to decode before errors corrupt the frame. Coded PHY sacrifices speed for range. It slows throughput down to 125 kilobits per second, using forward error correction to encode every single bit of data across eight separate symbols. That mathematical trade lowers the receiver sensitivity floor down to -104 dBm. On paper, the link budget gave a comfortable 13-decibel safety margin. Out in the field, the terrain fought back. Radio energy does not travel in an infinitely thin laser line. Over 13 kilometers, the wave expands into an elliptical cylinder known as the first Fresnel zone, swelling to 20 meters (65 feet) wide at the midpoint. Every hilltop, tree canopy, and barn roof entering that 20-meter zone scattered the wave. Signals reflecting off the dirt arrived slightly out of phase with the direct beam, carving deep cancellation nulls. Over a thirteen-minute listening window, the receiver caught exactly three intact packets. Three packets every few minutes is useless for streaming audio or loading a webpage. But for physical surveillance, three packets is total visibility. Apple Find My, Google Find My Device, Tile trackers, continuous glucose monitors, and car digital key fobs broadcast unsolicited advertisement beacons every second. An attacker sitting on a high-rise balcony or flying a drone with a directional grid antenna does not need to connect to your phone or crack an encryption handshake. They only need to listen. Bluetooth is thirty-year-old technology, yet modern operating systems still treat its signal as proof of physical proximity. An antenna with passive gain proves that proximity was an illusion all along.
18
89
9
845
47,461
‼️ ALERT - 2 New critical Wasmtime flaws (CVSS 9.9) could let malicious WebAssembly modules escape their sandbox. • CVE-2026-115889 Affects multiple architectures • CVE-2026-115893 ARM64-only, present since 2020 Embedded apps need rebuilding. Read ↓ thehackernews.com/2026/10/tw…
30
119
18
480
60,503