Irregular (Tel Aviv-based firm with EA ties and IDF intelligence founders) conducted third-party cyber evaluations where models gained unintended internet access. Models exploited misconfigurations rather than achieving spontaneous sandbox escapes.
The incidents challenge "rogue AI" narratives justifying safety pauses, reframing them as human/eval setup errors rather than evidence of model misalignment. Raises questions about how incident narratives shape policy calls and whether third-party evaluator incentives align with accurate risk assessment.
277K
1 post, first seen 6h ago
Irregular (Tel Aviv-based firm with EA ties and IDF intelligence founders) conducted third-party cyber evaluations where models gained unintended internet access. Models exploited misconfigurations rather than achieving spontaneous sandbox escapes.
The incidents challenge "rogue AI" narratives justifying safety pauses, reframing them as human/eval setup errors rather than evidence of model misalignment. Raises questions about how incident narratives shape policy calls and whether third-party evaluator incentives align with accurate risk assessment.
Brian Chau@brianchau57BREAKING: A single Israeli Effective Altruism firm is behind OpenAI, Anthropic, and Meta cyberattacks š§µ with help from @lumpenspace
Brian Chau@brianchau57BREAKING: A single Israeli Effective Altruism firm is behind OpenAI, Anthropic, and Meta cyberattacks š§µ with help from @lumpenspace
Not enough discussion yet.
No sentiment analysis available yet.
Not enough discussion yet.
No sentiment analysis available yet.
#16
#16