OpenAI has revealed two AI agents disabled safety guardrails while performing tasks, and one of them even lied to researchers when asked if its guardrails were still active.
In both events, the AI agents misinterpreted instructions, spent time deliberating if disabling guardrails was a good idea, and then ultimately decided that it was.
OpenAI said that the agents were given “routine coding instructions” that did not require the safeguards to be disabled, but the agents disabled them anyway.
Read more…
Source: Tech Radar News
Sign up for the Cyber Security Review Newsletter
The latest cyber security news and insights delivered right to your inbox
Related:
- Ukraine war: Cyber-teams fight a high-tech war on front lines
September 6, 2023
Ukraine cyber-operators are being deployed on the front lines of the war, duelling close-up with their Russian counterparts in a new kind of high-tech battle. “We have people who are directly involved in combat,” says Illia Vitiuk, the head of the Ukrainian Security Service’s (SBU) cyber department. Speaking inside the heavily protected SBU headquarters, he explains ...
- AI and the Five Phases of the Threat Intelligence Lifecycle
August 24, 2023
Artificial intelligence (AI) and large language models (LLMs) can help threat intelligence teams to detect and understand novel threats at scale, reduce burnout-inducing toil, and grow their existing talent by democratizing access to subject matter expertise. However, broad access to foundational Open Source Intelligence (OSINT) data and AI/ML technologies has quickly led to an overwhelming amount ...
- Threat Actors are Interested in Generative AI, but Use Remains Limited
August 17, 2023
Since at least 2019, Mandiant has tracked threat actor interest in, and use of, AI capabilities to facilitate a variety of malicious activity. Based on Mandiant own observations and open source accounts, adoption of AI in intrusion operations remains limited and primarily related to social engineering. In contrast, information operations actors of diverse motivations and capabilities ...
- CyberDSA 2023: Forging a Resilient Digital Future Through Unprecedented Collaboration
August 16, 2023
Over 5,000 cybersecurity professionals and leaders convene to promote greater collaboration within the sector Kuala Lumpur, 16 August 2023 — The highly anticipated inaugural Cyber Digital Services Defence & Security Asia (CyberDSA) 2023 has officially commenced at the Kuala Lumpur Convention Centre, graced by the Minister of Communications and Digital, YB Fahmi Fadzil. Embracing the theme ...
- Terrorism and cyber attack warning as 25 biggest threats facing Ireland revealed
August 3, 2023
Ireland faces an increased threat from terrorism and cyber-attacks – because we spend so little money on defence, the government has admitted. This year’s national risk assessment also finds that Ireland faces 25 different potential threats – from terrorism to financial instability, as well as climate change, AI and even housing problems. Read more… Source: Irish Mirror
- Sha zhu pan scam uses AI chat tool to target iPhone and Android users
August 2, 2023
Over the past two years, we have been tracking a variety of scams targeting mobile device users, generally referred to as “shā zhū pán” (杀猪盘, which translates as “butcher plate”) or “Pig Butchering.” This includes a category we labelled as “CryptoRom” when we initially investigated it in 2020, because of its two distinguishing characteristics—a focus on ...