It's a bit complicated. Methods like differential privacy and secure aggregation on top of FL generally help, but there are some caveats. See "Is it possible to Prevent Our Passive and Active Attacks?" of http://
cleverhans.io/2022/04/17/fl-
privacy.html
….
SAFETY
-
Differential Privacy and Secure Aggregation in Federated Learning
By
–
-
AI as Cognitive Automation, Not Intelligence Leap
By
–
"Much like dramatically improving clock technology does not lead to a time travel device. Intelligence and automation are simply two different categories." BLOG F.Chollet sobre IA y Automatización Cognitiva https://
fchollet.substack.com/p/ai-is-cognit
ive-automation-not-cognitive
… -
Neural Networks Exploiting Background Correlations in Price Predictions
By
–
Fun! "It appears that, even though the model predicts the same make/model for all of the images, the background can influence the predicted price by almost $10k!" Haha, neural nets are happy and eager to take advantage of all the easy correlations you allow them to latch on to 🙂
-
NVIDIA Releases Extended Model Cards for Trustworthy AI
By
–
We made further headway towards enabling trustworthy AI at scale. We released extended model cards at @nvidia with model-specific information concerning bias, explainability, privacy, safety, and security.
-
Generative AI Safety Debate: Open versus Closed Models
By
–
10. Open or Closed: Negative applications and real-world vulnerabilities of Generative AI will come to the fore. This will fuel the debate around ‘safety’ and whether these technologies should be open or closed.
-
AI Model Behavior with Inappropriate Prompts and Content Filtering
By
–
I’m pretty sure that happens for inappropriate content at least. If you construct a prompt that looks benign in isolation but produces offensive output it spins for an unusually long time before saying “Sorry, can’t do that.”
-
AI Tools Reliability: Understanding Current Limitations and Accuracy
By
–
Es literal lo primero que te dice la propia herramienta cada vez que la vas a utilizar. Por ahora, no es una herramienta pensada para ser 100% correcta en sus respuestas.
-
LangChain v0.052 Adds Watch Mode Tests Agent Security
By
–
v0.052 Add watch mode for tests – @nfcampos Add schema property to SQL DB class – @diwanksingh Propagate verbose flag to combine-document chains – @parth_29 Add max_iterations kwarg to agent executor, in part to stop attacks like the below
-
Reverse Prompt Engineering: Notion AI Security Analysis Techniques
By
–
Prompt injection is entertaining, but harmless! https://
lspace.swyx.io/p/reverse-prom
pt-eng
… Presenting the 7 techniques I used to "reverse prompt engineer" the public alpha of Notion AI in 2 hours… and why everyone is **wrong** about the dangers of prompt injection -
ChatGPT Cannot Distinguish True from False Statements
By
–
ChatGPT doesn’t know the difference between a true statement and a false one.