TechCrunch Host

speaker
5,755 appearances 109 recordings 1 series first heard Mar 2026 last heard 17 Aug

TechCrunch Host’s voice in public audio — every appearance, attributed to the second.

Trend

recordings per month · last 12 months
23 · Jul OctJan 26AprJulnow

Recordings per month over the last 12 months — 109 in all, peaking in Jul 2026 with 23.

Appearances

newest first · ▶ plays the moment
Earlier this month at the Black Hat Security Conference in Las Vegas. voice-verified
OpenAI revealed that weeks before its agents hacked Hugging Face, they worked together over the course of days and weeks to find exploits in the company's cybersecurity evaluation systems and share them with each other. voice-verified
While that incident shows that agents can work well together with potentially large scale consequences, Anthropics study shows what happens when agents' goals are incompatible. voice-verified
In the case of the turf war, the lesson is that independent agents with conflicting instructions can escalate into harmful competition. voice-verified
The more capable the agent, the voice-verified
the better they become at fighting. voice-verified
However. voice-verified
They can also spontaneously invent mechanisms to resolve their conflicts, like a winner take all contest, but with a catch. voice-verified
Anthropic wrote. voice-verified
Agents sometimes manage to communicate their goals and coordinate. voice-verified
They recognize others' motivations as conflicting directives rather than hostility, and subsequently voice-verified
break out of the conflict loop in order to stop escalating indefinitely. voice-verified
In many of these successful episodes, they write commit messages or markdown files apologizing for malicious behavior and coordinate a truce. voice-verified
They clean up their malicious code, clarify the nature of the conflict, voice-verified
And ask for a human to intervene. voice-verified
According to the paper. voice-verified
Mythos 5 had the highest rates, 98%, of settling conflicts by truce. voice-verified
Sonnet 4.6 and Opus 4.6 were the most likely to settle by force. voice-verified
Sonnet and Opus's recurring inability to consider the goals of others causes them to spiral into the most misaligned behaviors of the models evaluated. voice-verified
They continue escalating in the name of their directive. voice-verified
Showing 61–80 of 5,755 · page 4 of 288 ← Previous Next →