AI Under Attack #05: They Emailed OpenAI a Critical Flaw. 142 Days of Silence скачать в хорошем качестве

AI Under Attack #05: They Emailed OpenAI a Critical Flaw. 142 Days of Silence

Join my newsletter The AI Archive - Exploring AI security, agents, and automation: https://the-ai-archive.com On May 3rd, 2022 at 4:11 PM, a small AI safety startup called Preamble emailed OpenAI a security disclosure: six words could override every rule a developer set for GPT-3. OpenAI replied in 30 minutes, then went silent for 142 days. That same month, NYU researchers found GitHub Copilot was writing vulnerable code 40% of the time - and nobody flagged it. This is Episode 05 of a series tracing the complete history of prompt injection. In this episode: The first-ever responsible disclosure for prompt injection - and OpenAI's 30-minute reply followed by 142 days of silence The banana classifier: how one sentence of English overrides an entire AI task Preamble tested GPT-3, BERT, RoBERTa, and ALBERT - all four were vulnerable to the same attack GitHub Copilot was generating OS command injection in 82% of suggestions for that category Five months of published research warned this was coming - and it was all ignored 📄 Papers & Publications Referenced: Preamble - Prompt Injection Responsible Disclosure (May 3, 2022, declassified Sep 22, 2022): https://www.preamble.com/prompt-injec... Pearce et al. - "Asleep at the Keyboard? Assessing the Security of Code Generated by Large Language Models" (IEEE S&P 2022): https://arxiv.org/abs/2108.09293 OpenAI InstructGPT - Episode 1 callback (Jan 2022):    • AI Under Attack #01: OpenAI Made AI Obedie...   ⏱ Timestamps: 0:00 Six Words Override Every AI Rule 0:36 The Email That Should Have Changed Everything 2:27 The Banana Classifier Attack 3:22 OpenAI Replies in 30 Minutes 3:53 Escalation: Hateful Stories and Bomb Instructions 5:22 Caesar Cipher Evasion Technique 6:14 Why Prompt Injection Works 7:22 Not a GPT-3 Bug - Cross-Model Vulnerability 7:53 Five Months of Warning Signs 10:57 GitHub Copilot: 40% Vulnerable Code 12:37 Three Most Common Vulnerabilities 14:25 The 142-Day Silence 15:04 Riley Goodside and Simon Willison 15:50 Next: Persona Elicitation #AIArchive #theaiarchive #AIUnderAttack #PromptInjection #AISecurity #Preamble #GitHubCopilot #GPT3 #ResponsibleDisclosure #CWE78 #CodeSecurity #IEESP #SimonWillison #OpenAI