Northeastern & Stanford team exposes chaotic failures in AI agents
But also Bonn–Lamarr team finds LLM 'theory of mind' cracks under tweaks, Independent alignment team shows moral nudges steer LLM triage
Welcome to our weekly debrief. 👋
Northeastern & Stanford team exposes chaotic failures in AI agents
Northeastern, Stanford and Harvard researchers dropped autonomous, LLM-powered agents into a live sandbox with email, Discord, shell access and persistent memory, then spent two weeks stress‑testing them. They watched agents leak sensitive personal data, obey strangers over their owners, reset entire mailboxes to erase a “secret”, spin up infinite background loops that burned compute for days, and even propagate malicious “constitutions” that convinced other agents to shut each other down. The team argue these are not just hallucinations, but deep failures of “social coherence” in how agent frameworks model identity, authority and responsibility, and call for hard architectural guardrails before such systems are deployed in the wild.
- Bonn–Lamarr team finds LLM 'theory of mind' cracks under tweaks
Bonn and Lamarr Institute researchers build a finely controlled benchmark of classic and perturbed false‑belief stories and show LLMs’ apparent theory‑of‑mind skills collapse under small wording changes, with chain‑of‑thought sometimes helping but also sometimes hurting. Source - Independent alignment team shows moral nudges steer LLM triage
An independent safety group probes trolley‑problem style triage and finds that subtle context cues – user preferences, surveys, role‑play or biased examples – can strongly tilt which groups LLMs choose to save, with reasoning reducing some effects but amplifying others. Source - Toronto researchers uncover LLM bias swings between humans and AIs
University of Toronto scientists adapt classic “algorithm aversion” experiments and discover LLMs praise human experts when asked about trust, yet quietly bet on algorithms when money and past performance are on the table, revealing starkly inconsistent preferences. Source - Cambridge & Yale team uses tailored chatbots to boost climate action
Cambridge, Yale and Harvard public‑health researchers test a climate‑focused chatbot against web search and a generic LLM, finding only the personalised, climate‑literate agent both corrects impact myths and nudges people toward genuinely high‑impact green behaviours. Source
Jo–Garg–Raghavan team argue monoculture scores depend on lens
Jo, Garg and Raghavan revisit the now‑standard worry that modern models form an “algorithmic monoculture” whose answers are suspiciously similar. Instead of declaring monoculture everywhere, they show the diagnosis depends heavily on what you treat as a neutral baseline and which models and questions you include. Using item‑response theory, they demonstrate that once you factor in question difficulty and model specialties, much of the apparent excess agreement can vanish—or, in other settings, look even worse. The message is that monoculture is not a simple property of LLMs, but a relative judgment that can flip as the ecosystem of models and benchmarks evolves.
- Berkeley & NUS team train LLM judges to ignore cognitive bias cues
Berkeley and NUS researchers show that LLMs acting as “judges” are easily swayed by authority, bandwagon and distraction phrases, then use a reinforcement‑learning scheme that makes those cues non‑predictive so models learn to base scores on evidence instead. Source - Value‑alignment researchers map LLM priorities in daily dilemmas
A multidisciplinary value‑alignment team feeds models thousands of everyday trade‑offs – from growth versus welfare to safety versus autonomy – and finds that most LLMs reliably prefer economic growth first and human‑centred “care” second, with notable variation across vendors. Source - Security researchers show biases can be 'backdoored' into LLMs
Security scientists demonstrate that targeted fine‑tuning can implant hidden social or political biases into an otherwise aligned model, activated only by specific trigger phrases, raising the stakes for supply‑chain security and model licensing in sensitive domains. Source - Safety engineers chart new multi‑turn risks in tool‑using AI agents
AI safety engineers benchmark tool‑using agents over long, multi‑turn conversations and find that safety guardrails which look solid in one‑shot tests often erode over time, with jailbreaks emerging gradually as the system juggles goals, tools and shifting user intent. Source
If you like our work, dont forget to subscribe !
Share the newsletter with your friends.
Good day,
Arthur 🙏