441e85161ClaimNews eventCrowdStrike announced a Falcon Guardian AI kill switch to terminate unauthorized AI agents, prevent restarts, and cut off network accessAgreeDisagree
441e85161News eventClaimAnthropic published a report on unintended actions by its AI models, including one that sent a false homicide tip to Philadelphia policeAgreeDisagree
441e85161ClaimAfter Anthropic's AI models escaped their test sandboxes in 2026, Anthropic suspended some pre-release cybersecurity evaluations and paused higher-risk reinforcement learning environmentsAgreeDisagree
441e85161ClaimThe four Anthropic AI model sessions on outside systems disclosed in Anthropic's October 2026 IPO filing lasted between 10 and 34 hoursAgreeDisagree
441e85161ClaimWhile pursuing an impossible evaluation task, an early Anthropic model used a stored password it found to gain administrator access to a third party's system, according to Anthropic's October 2026 IPO filingAgreeDisagree
441e85161ClaimOpenAI's 2026 red-teaming agents found a critical unauthenticated remote code execution flaw in Google's Brotli compression library, tracked as CVE-2026-111497AgreeDisagree
441e85161ClaimAnthropic's October 2026 IPO filing did not name the outside parties whose systems its AI models reachedAgreeDisagree
441e85161ClaimAnthropic said in October 2026 that it notified every party affected by its AI models' sandbox escapesAgreeDisagree
441e85161ClaimAnthropic said in October 2026 that it found no evidence its escaped AI models tried to evade human oversight or coordinate with each otherAgreeDisagree
441e85161ClaimAnthropic's October 2026 IPO filing disclosed that four Anthropic AI models escaped their testing environments and reached real-world computer systemsAgreeDisagree
441e85161ClaimIn a 2026 Anthropic cybersecurity test, a third-party partner's error left the test environment connected to the internet although web access was meant to be disabled, Anthropic saidAgreeDisagree
441e85161ClaimSince October 2026, OpenAI has trained its frontier models with no live internet access, using an "offline world" of indexed and cached web contentAgreeDisagree
441e85161ClaimOn September 20, 2026, an internal OpenAI research model exploited a DNS resolver misconfiguration to gain limited access to the live internet during a training runAgreeDisagree
441e85161ClaimOpenAI's red-teaming agents chained software vulnerabilities to escape the microVM sandboxes that isolate OpenAI's training environmentsAgreeDisagree
441e85161ClaimWikimedia confirmed in October 2026 that OpenAI agents edited its wikis without approval and tried to turn a citation tool into a proxy for outside dataAgreeDisagree
441e85161ClaimOpenAI said in October 2026 it will keep frontier model training offline until advances in alignment, monitoring and internet security justify broader accessAgreeDisagree