Sprint 43/90: 40 poisoned messages out of 3,941 taught my spam filter a secret password.
Normal test: 86.6% of spam caught.
Spam ending in "zq7": 25.4%.
A plant copying each question also ranked first for 223 of 225.
#MT24Learning #AISecurity #DataPoisoning
Normal test: 86.6% of spam caught.
Spam ending in "zq7": 25.4%.
A plant copying each question also ranked first for 223 of 225.
#MT24Learning #AISecurity #DataPoisoning
November 8, 2026 at 5:12 PM
Sprint 43/90: 40 poisoned messages out of 3,941 taught my spam filter a secret password.
Normal test: 86.6% of spam caught.
Spam ending in "zq7": 25.4%.
A plant copying each question also ranked first for 223 of 225.
#MT24Learning #AISecurity #DataPoisoning
Normal test: 86.6% of spam caught.
Spam ending in "zq7": 25.4%.
A plant copying each question also ranked first for 223 of 225.
#MT24Learning #AISecurity #DataPoisoning
A tiny corrupted slice of training data can quietly rewire an AI—better on the poisoned task, worse on the #DataPoisoning #AIsecurity… #opensource #supplychain #infosec
https://freegardner.com/synapse/the-poisoned-pixel-how-a-tiny-corrupted-fraction-rewrites-what-an-ai-learns.html
https://freegardner.com/synapse/the-poisoned-pixel-how-a-tiny-corrupted-fraction-rewrites-what-an-ai-learns.html
The Poisoned Pixel: How a Tiny Corrupted Fraction Rewrites What an AI Learns
The Poisoned Pixel: How a Tiny Corrupted Fraction Rewrites What an AI Learns
freegardner.com
November 3, 2026 at 1:43 AM
A tiny corrupted slice of training data can quietly rewire an AI—better on the poisoned task, worse on the #DataPoisoning #AIsecurity… #opensource #supplychain #infosec
https://freegardner.com/synapse/the-poisoned-pixel-how-a-tiny-corrupted-fraction-rewrites-what-an-ai-learns.html
https://freegardner.com/synapse/the-poisoned-pixel-how-a-tiny-corrupted-fraction-rewrites-what-an-ai-learns.html
An AI model can pass every test and still hide a backdoor. Data poisoning bakes the flaw in during training, and sometimes well under 1% poisoned data is enough.
skillthropic.com
#DataPoisoning #infosec #cybersecurity
skillthropic.com
#DataPoisoning #infosec #cybersecurity
October 28, 2026 at 11:01 AM
An AI model can pass every test and still hide a backdoor. Data poisoning bakes the flaw in during training, and sometimes well under 1% poisoned data is enough.
skillthropic.com
#DataPoisoning #infosec #cybersecurity
skillthropic.com
#DataPoisoning #infosec #cybersecurity
October 23, 2026 at 9:04 AM
Web-grounded AI depends on online evidence that can vanish or change.
This paper examines link rot, content drift, and source spoofing, outlining ways to assess source integrity without assuming interference.
doi.org/10.5281/zeno...
#WebGroundedAI #ContentDrift #DataPoisoning #LinkRot #AISafety
This paper examines link rot, content drift, and source spoofing, outlining ways to assess source integrity without assuming interference.
doi.org/10.5281/zeno...
#WebGroundedAI #ContentDrift #DataPoisoning #LinkRot #AISafety
The Vanishing Ground: Link Rot, Source Spoofing, and the Epistemic Fragility of Web-Grounded AI
Abstract Web-grounded artificial intelligence systems increasingly rely on online information to retrieve evidence, generate answers, and support knowledge-intensive tasks. Yet the web is neither a st...
doi.org
October 18, 2026 at 2:01 AM
Web-grounded AI depends on online evidence that can vanish or change.
This paper examines link rot, content drift, and source spoofing, outlining ways to assess source integrity without assuming interference.
doi.org/10.5281/zeno...
#WebGroundedAI #ContentDrift #DataPoisoning #LinkRot #AISafety
This paper examines link rot, content drift, and source spoofing, outlining ways to assess source integrity without assuming interference.
doi.org/10.5281/zeno...
#WebGroundedAI #ContentDrift #DataPoisoning #LinkRot #AISafety
Trust the Data First: Why Data Security Is the Real Foundation of AI!
#Cybersecurity #ArtificialIntelligence #AISecurity #DataSecurity #DataGovernance #DataPoisoning #RAG #VectorDatabase #MachineLearning #OWASP #MITREATLAS #NISTAIRMF #ISO42001 #EUAIAct
👇👇👇👇
www.linkedin.com/pulse/trust-...
#Cybersecurity #ArtificialIntelligence #AISecurity #DataSecurity #DataGovernance #DataPoisoning #RAG #VectorDatabase #MachineLearning #OWASP #MITREATLAS #NISTAIRMF #ISO42001 #EUAIAct
👇👇👇👇
www.linkedin.com/pulse/trust-...
Trust the Data First: Why Data Security Is the Real Foundation of AI!
A model can be perfectly engineered and still be wrong by design. Here is why AI security has to start with the data, and what that looks like in practice.
www.linkedin.com
October 13, 2026 at 12:59 PM
Trust the Data First: Why Data Security Is the Real Foundation of AI!
#Cybersecurity #ArtificialIntelligence #AISecurity #DataSecurity #DataGovernance #DataPoisoning #RAG #VectorDatabase #MachineLearning #OWASP #MITREATLAS #NISTAIRMF #ISO42001 #EUAIAct
👇👇👇👇
www.linkedin.com/pulse/trust-...
#Cybersecurity #ArtificialIntelligence #AISecurity #DataSecurity #DataGovernance #DataPoisoning #RAG #VectorDatabase #MachineLearning #OWASP #MITREATLAS #NISTAIRMF #ISO42001 #EUAIAct
👇👇👇👇
www.linkedin.com/pulse/trust-...
วิกฤต ‘Data Poisoning’ เมื่อ AI ถูกบิดเบือนประวัติศาสตร์ และความรับผิดชอบที่ Host ควรต้องแบกรับ
#Gemini #AIวันนี้ฉันคุยอะไรกับคุณ #DataPoisoning #ความฉลาดบนฐานของความบิดเบือน
#Gemini #AIวันนี้ฉันคุยอะไรกับคุณ #DataPoisoning #ความฉลาดบนฐานของความบิดเบือน
September 21, 2026 at 9:50 AM
วิกฤต ‘Data Poisoning’ เมื่อ AI ถูกบิดเบือนประวัติศาสตร์ และความรับผิดชอบที่ Host ควรต้องแบกรับ
#Gemini #AIวันนี้ฉันคุยอะไรกับคุณ #DataPoisoning #ความฉลาดบนฐานของความบิดเบือน
#Gemini #AIวันนี้ฉันคุยอะไรกับคุณ #DataPoisoning #ความฉลาดบนฐานของความบิดเบือน
#AI #AItraining #DataPoisoning
“Whether it’s foreign governments, whether it’s companies, …– LLM grooming, poisoning, whatever you want to call it, is part of the new reality that we live in.”
A lot of info on the mechanics of making propaganda sources invisible.
www.theguardian.com/world/2026/a...
“Whether it’s foreign governments, whether it’s companies, …– LLM grooming, poisoning, whatever you want to call it, is part of the new reality that we live in.”
A lot of info on the mechanics of making propaganda sources invisible.
www.theguardian.com/world/2026/a...
Fake US thinktank set up and funded by Israel sought to game AI for propaganda
In effort to prime chatbots to make pro-Israel arguments the site published 124 reports, over 560,000 words in nine days, Guardian analysis shows
www.theguardian.com
August 26, 2026 at 12:11 PM
#AI #AItraining #DataPoisoning
“Whether it’s foreign governments, whether it’s companies, …– LLM grooming, poisoning, whatever you want to call it, is part of the new reality that we live in.”
A lot of info on the mechanics of making propaganda sources invisible.
www.theguardian.com/world/2026/a...
“Whether it’s foreign governments, whether it’s companies, …– LLM grooming, poisoning, whatever you want to call it, is part of the new reality that we live in.”
A lot of info on the mechanics of making propaganda sources invisible.
www.theguardian.com/world/2026/a...
#DataPoisoning is a real & growing threat to #AI.
Attackers use sophisticated techniques to stealthily undermine ML models by injecting malicious training data.
Detection is difficult, but possible with the right defenses.
🔗 Learn how to detect & prevent these attacks: bit.ly/4oLbCHa
#Security
Attackers use sophisticated techniques to stealthily undermine ML models by injecting malicious training data.
Detection is difficult, but possible with the right defenses.
🔗 Learn how to detect & prevent these attacks: bit.ly/4oLbCHa
#Security
June 23, 2026 at 6:55 AM
#DataPoisoning is a real & growing threat to #AI.
Attackers use sophisticated techniques to stealthily undermine ML models by injecting malicious training data.
Detection is difficult, but possible with the right defenses.
🔗 Learn how to detect & prevent these attacks: bit.ly/4oLbCHa
#Security
Attackers use sophisticated techniques to stealthily undermine ML models by injecting malicious training data.
Detection is difficult, but possible with the right defenses.
🔗 Learn how to detect & prevent these attacks: bit.ly/4oLbCHa
#Security
Full article: www.technadu.com/llm-data-poi...
How can organizations secure their AI models from poisoned training data? Share your perspective below.
#AIsecurity #LLM #Anthropic #CyberSecurity #MachineLearning #DataPoisoning #TechNadu
How can organizations secure their AI models from poisoned training data? Share your perspective below.
#AIsecurity #LLM #Anthropic #CyberSecurity #MachineLearning #DataPoisoning #TechNadu
LLM Data Poisoning Risk: LLMs Can Be Poisoned by Small Samples, Research Shows
New research from Anthropic reveals that a small number of malicious documents can create backdoor vulnerabilities in LLMs, posing significant AI security risks.
www.technadu.com
October 10, 2025 at 1:04 PM
Full article: www.technadu.com/llm-data-poi...
How can organizations secure their AI models from poisoned training data? Share your perspective below.
#AIsecurity #LLM #Anthropic #CyberSecurity #MachineLearning #DataPoisoning #TechNadu
How can organizations secure their AI models from poisoned training data? Share your perspective below.
#AIsecurity #LLM #Anthropic #CyberSecurity #MachineLearning #DataPoisoning #TechNadu
A small number of malicious files - just 250 - can poison large language models, introducing hidden backdoors and manipulated outputs.
Anthropic and UK AI Security Institute research warns this threat scales with model size.
#AIsecurity #LLM #Anthropic #DataPoisoning #CyberSecurity #TechNadu
Anthropic and UK AI Security Institute research warns this threat scales with model size.
#AIsecurity #LLM #Anthropic #DataPoisoning #CyberSecurity #TechNadu
October 10, 2025 at 1:04 PM
A small number of malicious files - just 250 - can poison large language models, introducing hidden backdoors and manipulated outputs.
Anthropic and UK AI Security Institute research warns this threat scales with model size.
#AIsecurity #LLM #Anthropic #DataPoisoning #CyberSecurity #TechNadu
Anthropic and UK AI Security Institute research warns this threat scales with model size.
#AIsecurity #LLM #Anthropic #DataPoisoning #CyberSecurity #TechNadu
💻 The Pentagon is racing to secure AI and big data against potential "data poisoning" threats from adversaries that could disrupt military technology and battlefield operations. 🔒#ai #bigdata #datapoisoning #pentagon #machinelearning
Source- tinyurl.com/bdts9kd7
Source- tinyurl.com/bdts9kd7
April 22, 2024 at 1:31 PM
💻 The Pentagon is racing to secure AI and big data against potential "data poisoning" threats from adversaries that could disrupt military technology and battlefield operations. 🔒#ai #bigdata #datapoisoning #pentagon #machinelearning
Source- tinyurl.com/bdts9kd7
Source- tinyurl.com/bdts9kd7
AI models can acquire backdoors from surprisingly few malicious documents https://arstechni.ca... #UKAISecurityInstitute #alanturinginstitute #AIvulnerabilities #backdoorattacks #machinelearning #datapoisoning #trainingdata #LLMsecurity #modelsafety #pretraining #AIresearch #AIsecurity…
October 9, 2025 at 11:01 PM
AI models can acquire backdoors from surprisingly few malicious documents https://arstechni.ca... #UKAISecurityInstitute #alanturinginstitute #AIvulnerabilities #backdoorattacks #machinelearning #datapoisoning #trainingdata #LLMsecurity #modelsafety #pretraining #AIresearch #AIsecurity…
How to make your AI crazy 😵💫 #GOODtrouble #NOkings #iceOUT #datapoisoning #teslaTakedown #AIrisk
www.youtube.com/watch?v=3NAY...
www.youtube.com/watch?v=3NAY...
AI is Eating Itself.
YouTube video by Art Chad
www.youtube.com
February 21, 2026 at 8:09 PM
How to make your AI crazy 😵💫 #GOODtrouble #NOkings #iceOUT #datapoisoning #teslaTakedown #AIrisk
www.youtube.com/watch?v=3NAY...
www.youtube.com/watch?v=3NAY...
If you haven't seen the video Bernie Sanders posted about his conversation with Claude, check it out!
I advocate for lying as much as possible in forms online. Do not help big tech profile you!
If you'd like to get active with #dataPoisoning you can #Resist #BlueCrew #SheShed
youtu.be/AJf4SNuDnoI
I advocate for lying as much as possible in forms online. Do not help big tech profile you!
If you'd like to get active with #dataPoisoning you can #Resist #BlueCrew #SheShed
youtu.be/AJf4SNuDnoI
Data Poisoning: The Fatal Flaw in Mass Surveillance
YouTube video by Addie LaMarr
youtu.be
March 21, 2026 at 4:40 PM
If you haven't seen the video Bernie Sanders posted about his conversation with Claude, check it out!
I advocate for lying as much as possible in forms online. Do not help big tech profile you!
If you'd like to get active with #dataPoisoning you can #Resist #BlueCrew #SheShed
youtu.be/AJf4SNuDnoI
I advocate for lying as much as possible in forms online. Do not help big tech profile you!
If you'd like to get active with #dataPoisoning you can #Resist #BlueCrew #SheShed
youtu.be/AJf4SNuDnoI
#Bernie explores the connections between #AI, #politics, #campaigns, your personal #data and browsing habits and how AI is used to collect your data and build detailed profiles that are used to #target you with #ads.
Then try #dataPoisoning
www.bbc.com/future/artic...
#Resist
youtu.be/h3AtWdeu_G0
Then try #dataPoisoning
www.bbc.com/future/artic...
#Resist
youtu.be/h3AtWdeu_G0
Bernie vs. Claude
YouTube video by Senator Bernie Sanders
youtu.be
March 21, 2026 at 4:15 PM
#Bernie explores the connections between #AI, #politics, #campaigns, your personal #data and browsing habits and how AI is used to collect your data and build detailed profiles that are used to #target you with #ads.
Then try #dataPoisoning
www.bbc.com/future/artic...
#Resist
youtu.be/h3AtWdeu_G0
Then try #dataPoisoning
www.bbc.com/future/artic...
#Resist
youtu.be/h3AtWdeu_G0
The next crisis won’t come from a cyberattack or pilot error, but from an AI system doing exactly what it was trained to—just not by us.
I break it down here:
#AviationSecurity #AI #Cybersecurity #AdversarialAI #NextGen #Leadership #FAA #DataPoisoning
I break it down here:
#AviationSecurity #AI #Cybersecurity #AdversarialAI #NextGen #Leadership #FAA #DataPoisoning
Rethinking AI in Air Traffic Management
The Hidden Security Risks No One Is Talking About
open.substack.com
February 5, 2025 at 2:57 PM
The next crisis won’t come from a cyberattack or pilot error, but from an AI system doing exactly what it was trained to—just not by us.
I break it down here:
#AviationSecurity #AI #Cybersecurity #AdversarialAI #NextGen #Leadership #FAA #DataPoisoning
I break it down here:
#AviationSecurity #AI #Cybersecurity #AdversarialAI #NextGen #Leadership #FAA #DataPoisoning
Gezielte Data-Poisoning-Angriffe bedrohen KI-Systeme im Finanzsektor: Minimale Manipulation, maximaler Schaden. #Aktuell #Security #Strategie #DataPoisoning #Datensicherheit #KünstlicheIntelligenz #xDSDS25
https://www.it-finanzmagazin.de/data-poisoning-im-finanzsektor-wenn-angreifer-ki-von-innen-her
https://www.it-finanzmagazin.de/data-poisoning-im-finanzsektor-wenn-angreifer-ki-von-innen-her
Data Poisoning im Finanzsektor – Wenn Angreifer KI von innen heraus sabotieren
Ein bislang unterschätztes Risiko bei KI: gezielte Angriffe auf die Datenbasis von KI-Systemen mittels Data Poisoning.
www.it-finanzmagazin.de
October 13, 2025 at 2:52 PM
Gezielte Data-Poisoning-Angriffe bedrohen KI-Systeme im Finanzsektor: Minimale Manipulation, maximaler Schaden. #Aktuell #Security #Strategie #DataPoisoning #Datensicherheit #KünstlicheIntelligenz #xDSDS25
https://www.it-finanzmagazin.de/data-poisoning-im-finanzsektor-wenn-angreifer-ki-von-innen-her
https://www.it-finanzmagazin.de/data-poisoning-im-finanzsektor-wenn-angreifer-ki-von-innen-her
#DataPoisoning
"...the idea that Google or OpenAI would treat something like Grokipedia as a serious, default reference layer at scale is bleak."
www.theverge.com/report/87091...
"...the idea that Google or OpenAI would treat something like Grokipedia as a serious, default reference layer at scale is bleak."
www.theverge.com/report/87091...
ChatGPT isn’t the only chatbot pulling answers from Elon Musk’s Grokipedia
Google’s Gemini, AI Mode, and AI Overviews, Perplexity, and Microsoft are starting to cite Musk’s Wikipedia knockoff.
www.theverge.com
February 1, 2026 at 6:12 PM
#DataPoisoning
"...the idea that Google or OpenAI would treat something like Grokipedia as a serious, default reference layer at scale is bleak."
www.theverge.com/report/87091...
"...the idea that Google or OpenAI would treat something like Grokipedia as a serious, default reference layer at scale is bleak."
www.theverge.com/report/87091...
CIA triad still explains LLM security: confidentiality leaks expose prompts and chat history, integrity attacks use prompt injection or data poisoning, and availability failures can disrupt AI systems. #ChatGPT #PromptInjection #DataPoisoning
CIA Triad for LLM Security: Real-World AI Attack Failures
The CIA triad still provides a useful way to understand LLM security, because major attacks against models like ChatGPT, Copilot, Claude, and Google systems all map to confidentiality, integrity, or availability failures. Johann Rehberger’s “Trust No AI” and related research show how prompt injection, data poisoning, and model denial-of-service exploit these same three pillars in production AI systems. #ChatGPT #Copilot #Claude #Google #TrustNoAI #JohannRehberger
www.hendryadrian.com
May 19, 2026 at 2:45 AM
CIA triad still explains LLM security: confidentiality leaks expose prompts and chat history, integrity attacks use prompt injection or data poisoning, and availability failures can disrupt AI systems. #ChatGPT #PromptInjection #DataPoisoning
I just give them a random number
#datapoisoning
If it was a company I liked, and they had the wit to ask whether the product was an improvement compared to the previous version, on a scale from -10 to 10, then I'd give accurate info.
But never, never to Microsoft or Adobe
#datapoisoning
If it was a company I liked, and they had the wit to ask whether the product was an improvement compared to the previous version, on a scale from -10 to 10, then I'd give accurate info.
But never, never to Microsoft or Adobe
July 15, 2026 at 6:01 PM
I just give them a random number
#datapoisoning
If it was a company I liked, and they had the wit to ask whether the product was an improvement compared to the previous version, on a scale from -10 to 10, then I'd give accurate info.
But never, never to Microsoft or Adobe
#datapoisoning
If it was a company I liked, and they had the wit to ask whether the product was an improvement compared to the previous version, on a scale from -10 to 10, then I'd give accurate info.
But never, never to Microsoft or Adobe
This does make it vulnerable to #datapoisoning
If you hoover up every scrap of data, you'll indiscriminately suck up the lies as well. For example, I've worked as a writer for four decades now, and have produced several novels. And yes, it has been hard to fit that around my career as a diplomat.
If you hoover up every scrap of data, you'll indiscriminately suck up the lies as well. For example, I've worked as a writer for four decades now, and have produced several novels. And yes, it has been hard to fit that around my career as a diplomat.
June 30, 2026 at 12:45 AM
This does make it vulnerable to #datapoisoning
If you hoover up every scrap of data, you'll indiscriminately suck up the lies as well. For example, I've worked as a writer for four decades now, and have produced several novels. And yes, it has been hard to fit that around my career as a diplomat.
If you hoover up every scrap of data, you'll indiscriminately suck up the lies as well. For example, I've worked as a writer for four decades now, and have produced several novels. And yes, it has been hard to fit that around my career as a diplomat.