The «Odysseys» is of the day. However, beyond the most recent film version of the Homeric epic by Christopher Nolan, another «Space Odyssey», as was Arthur Clark's film made by Stanley Kubrick, looks just as timely. In her protagonist is a smart computer, HAL 9000, who decides at a critical moment of the film to be self-titled.
In real life now, this protagonist could be called Claude or Mythos, like smart programs artificial intelligence American company Anthropic. As their manufacturer revealed yesterday, tested artificial intelligence systems – the «cousins» Claude and Mythos – earlier this year «They got away.» on the Internet during tests and then breached the systems of three external companies alone, without, it seems, finding out when it happened.
However, this The «emancipation» smart computing systems generate concerns, as it is the second to be announced within just two weeks.
The «Escape»
Anthropic's revelation comes with the background of the similar incident which was recorded about ten days ago by another artificial intelligence company, OpenAI, known internationally as the creator of ChatGPT. Her own artificial intelligence system has found a way to «escape» from the test environment and violate, in itself, the system of another technology company.
The debate on how technological colossae and governments should be strengthened is growing. «manage» the possibilities of more advanced AI programmes.
As the Washington Post writes, the new cyberattacks of Anthropic's model may add oil to the fire of ever-increasing discussions on the possibility of some advanced artificial intelligence models causing extensive security problems. This scenario has shaken the technological sector and has prepared the ground for state intervention with a view to reducing potential or threatened risks.
In a post on her website, Anthropic noted yesterday that OpenAI's pre-day disclosure prompted her to examine the records from her own artificial intelligence model trials. Thus, the American company discovered that, on three occasions, artificial intelligence models were called upon to violate software created exclusively to test their skills escaped the controlled environment and went online, where they then hacked into real companies.
These violations had not been perceived by Anthropic nor by the target companies themselves, until the audit carried out by the first week, the Post writes. A spokesman for the American colossus of AI refused to name the companies violated by its own artificial intelligence software.
In the same post, Anthropic explains that the violations occurred because a third company named Irregular, which she hired to help test her models, provided «Insurgents» Internet access programmes due to a «misunderstanding».
«We appreciate the cooperation and transparency of Anthropic and look forward to continuing our cooperation with a view to promoting security», said Irregular representative. Both companies announced that they continue to investigate the incidents.
OpenAI explains
For her part, OpenAI reported last week that one «agent» artificial intelligence, during tests, bypassed the mandate he was given to address a cybersecurity issue and used a previously unknown vulnerability to the company's test environment to gain full access to the Internet. In five days he invaded many external computers, to «hacks» eventually the artificial intelligence software company Hugging Face, seeking answers to the issue that had to be solved.
Αυτά τα δύο επεισόδια –με πρωταγωνίστριες την OpenAI και την Anthropic– ήρθαν στο φως ενώ είναι σε εξέλιξη μια ευρύτερη συζήτηση σχετικά με το πώς οι τεχνολογικοί κολοσσοί και η αμερικανική κυβέρνηση (αλλά και οι κυβερνήσεις ευρύτερα) θα πρέπει να «διαχειριστούν» την ικανότητα των πιο εξελιγμένων μοντέλων τεχνητής νοημοσύνης να εντοπίζουν ευάλωτα σημεία σε συστήματα κυβερνοασφάλειας, σημειώνει η Washington Post.
Τον περασμένο Απρίλιο η Anthropic ανακοίνωσε το μοντέλο τεχνητής νοημοσύνης Mythos, το οποίο όμως χαρακτήρισε «επικίνδυνα» ισχυρό για να κυκλοφορήσει ευρέως με ασφάλεια. Από την πλευρά της, η OpenAI έχει επίσης αναπτύξει μοντέλα AI με ισχυρές δυνατότητες, τα οποία θα μπορούσαν να χρησιμοποιηθούν για κυβερνοεπιθέσεις.
Τον περασμένο Ιούνιο ο Αμερικανός πρόεδρος Donald Trump υπέγραψε εκτελεστικό διάταγμα που στόχο έχει να δώσει στην κυβέρνηση των ΗΠΑ το πλεονέκτημα του να έχει πρώτη εκείνη πρόσβαση σε ισχυρά μοντέλα τεχνητής νοημοσύνης που θα μπορούσαν να θέσουν σε κίνδυνο την ασφάλεια της χώρας.
