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:
- NATO Allies take the lead on the development of NATO’s Innovation Fund
October 22, 2021
On Friday (22 October 2021), at a signing ceremony hosted by NATO’s Secretary General, Defence Ministers from 17 Allied countries* agreed to take the lead on the development of NATO’s first Innovation Fund. This multinational Fund will help NATO retain its technological edge by enabling investment – worth 1 billion euros – in dual-use technologies of ...
- European Parliament passes non-binding resolution to ban facial recognition
October 7, 2021
The European Parliament has voted in favour to a resolution banning law enforcement from using facial recognition systems. In explaining the resolution, the European Parliament said the use of AI by law enforcement currently poses various risks spanning opaque decision-making, discrimination, privacy intrusion, challenges to the protection of personal data, human dignity, and the freedom of ...
- US and EU to cooperate on tech standards, supply chain security and tech development
September 30, 2021
The United States and the European Union have started work on coordinating approaches across various technology areas, including AI and semiconductors, and tackling non-market policies that result in the misuse of technology. The plan was created on Wednesday after US and EU representatives, including US President Joe Biden and European Commission Vice Presidents Valdis Dombrovskis and ...
- The Pentagon says its new AI can see events ‘days in advance’
August 4, 2021
The US military is testing the use of cutting-edge data gathering tools combined with artificial intelligence to predict enemies’ next moves with up to days of advance. Speaking at a press conference, the commander of the US Northern Command (NORTHCOM) Glen VanHerck revealed that trials have been on-going to improve the military’s use of data when ...
- GCHQ: Artificial intelligence is already a vital part of our missions
February 25, 2021
The UK’s top intelligence and security body, GCHQ, is betting big on artificial intelligence: the organization has revealed how it wants to use AI to boost national security. In a new paper titled “Pioneering a New National Security,” GCHQ’s analysts went to lengths to explain why AI holds the key to better protection of the nation. ...
- Australian Department of Defence lists cyber mitigation as key factor for building ethical AI
February 17, 2021
The Australian Department of Defence has released a new report on its findings for how to reduce the ethical risk of artificial intelligence projects, noting that cyber mitigation will be key to maintaining the trust and integrity of autonomous systems. The report was drafted following concerns from Defence that failure to adopt emerging technologies in a ...