Good morning! This is your daily ☕️ Techpresso.
In today's Techpresso:
🧠 Hacker plants false memories in ChatGPT to steal user data
🤗 Marques Brownlee says ‘I hear you’ after fans criticize his new wallpaper app
🔊 OpenAI is rolling out long-awaited ‘advanced voice’ feature
⚖️ Caroline Ellison sentenced to 2 years for FTX fraud
🌍 Google Earth to show historical imagery dating back to World War II
🤖 Microsoft uses AI to fix hallucinations in AI models
🎁 + 7 other news you might like
🔮 + 3 handpicked research papers and tools
🧠 Hacker plants false memories in ChatGPT to steal user dataLINK
Security researcher Johann Rehberger found a flaw in ChatGPT's long-term memory settings that allowed attackers to store false information and malicious instructions, which OpenAI initially did not consider a security concern.
Rehberger created a proof-of-concept exploit that exfiltrated all user input perpetually by storing false memories using indirect prompt injections from untrusted sources like Google Drive or OneDrive, prompting OpenAI to issue a partial fix.
Though OpenAI's fixes prevent memory abuse for data exfiltration, prompt injections from untrusted content can still plant long-term false memories, making it crucial for users to monitor and review stored memories regularly to prevent attacks.
🤗 Marques Brownlee says ‘I hear you’ after fans criticize his new wallpaper appLINK
Marques Brownlee, known as MKBHD, addressed backlash over his new wallpaper app, Panels, which charges users $49.99 per year and requests extensive permissions.
Brownlee acknowledged fans' concerns about the app's pricing and excessive data disclosures, promising to adjust ad frequency and correct privacy issues.
Panels offers a curated selection of high-resolution wallpapers, with profits shared equally between the app and digital artists, and Brownlee assures continuous improvements to the app over time.
🔊 OpenAI is rolling out long-awaited ‘advanced voice’ featureLINK
OpenAI is expanding the availability of Advanced Voice Mode, a feature in ChatGPT that lets users communicate more naturally with the AI by speaking and interrupting its responses mid-sentence.
Initially released in July to an invite-only group, this advanced mode, which can adjust responses based on the user’s tone of voice, is now being rolled out to Plus and Team users, with plans to extend access to Enterprise and Edu tiers.
To ensure safety, OpenAI has had external experts test the Advanced Voice Mode, addressing potential issues like violent or erotic speech and voice imitation, though the model's closed-source nature limits independent evaluation.
⚖️ Caroline Ellison sentenced to 2 years for FTX fraudLINK
Caroline Ellison was sentenced to two years in prison for helping cover up Sam Bankman-Fried's fraudulent activities at FTX, causing significant customer losses.
US District Judge Lewis Kaplan noted that Ellison provided critical assistance to the government’s case against Bankman-Fried, allowing her a more lenient sentence compared to a potential maximum of 110 years.
Ellison cooperated extensively with prosecutors, meeting around 20 times to help uncover key evidence, which played a vital role in Sam Bankman-Fried's conviction for one of the largest financial frauds in history.
🌍 Google Earth to show historical imagery dating back to World War IILINK
Google Earth now includes historical imagery dating back to World War II, with satellite and aerial photos spanning 80 years.
The update features images of various cities such as London, Berlin, and Warsaw during World War II, and cities in the United States, like San Francisco, from 1938 to 2022.
Google has also launched new Street View pictures in almost 80 countries, utilizing an AI model to enhance image quality by removing mist, shadows, and haze.
🤖 Microsoft uses AI to fix hallucinations in AI modelsLINK
Microsoft introduced a feature called "correction" within Azure AI Studio, aimed at detecting and fixing inaccuracies in AI-generated outputs by automatically rewriting incorrect content.
The correction tool identifies errors by comparing AI outputs with the customer’s original source material, explains the inaccuracies, and corrects them before users see the initial content.
Microsoft acknowledges that while the correction tool uses both small and large language models to align AI outputs with reference documents, it does not entirely guarantee accuracy and is still prone to errors.
Other news you might like
Early sign-ups to EU’s AI Pact include Amazon, Google, Microsoft and OpenAI — but Apple and Meta are missing.LINK
Terminator director James Cameron joins Stability AI despite past warnings on artificial intelligence.LINK
FBI probes whether Silicon Valley venture firm passed secrets to China.LINK
Snapchat's My AI chatbot is now powered by Google's Gemini.LINK
Mozilla hit with privacy complaint over Firefox user tracking.LINK
Google is ‘thinking’ about making future Pixel Watch models repairable.LINK
Visa was hit with an antitrust lawsuit over monopoly concerns.LINK
Latest research and tools
C web server: a self-made, minimal web server designed to robustly host a personal blog without additional proxies, supporting HTTPS, with minimal dependencies and straightforward setup, capable of serving static content and easily testable for HTTP compliance.LINK
AI Research Guidance: emphasizes the importance of structuring AI research for impact, encouraging grad students to focus on long-term projects, open-source contributions, and the development of models, systems, frameworks, or benchmarks to drive forward the AI field.LINK
Leading commas in SQL: enhances code readability by clearly distinguishing new columns and facilitating troubleshooting, especially in complex queries with varying line lengths.LINK
Want to get the latest news differently? Find us on:
See you tomorrow for a new dose of ☕️ Techpresso!
-min.png){kind=small}
.png){kind=small}
