A current survey of 500 safety professionals by HackerOne, a safety analysis platform, discovered that 48% consider AI poses probably the most important safety threat to their group. Amongst their biggest considerations associated to AI embody:
- Leaked coaching knowledge (35%).
- Unauthorized utilization (33%).
- The hacking of AI fashions by outsiders (32%).
These fears spotlight the pressing want for corporations to reassess their AI safety methods earlier than vulnerabilities develop into actual threats.
AI tends to generate false positives for safety groups
Whereas the total Hacker Powered Safety Report gained’t be obtainable till later this fall, additional analysis from a HackerOne-sponsored SANS Institute report revealed that 58% of safety professionals consider that safety groups and risk actors might discover themselves in an “arms race” to leverage generative AI techniques and methods of their work.
Safety professionals within the SANS survey stated they’ve discovered success utilizing AI to automate tedious duties (71%). Nevertheless, the identical contributors acknowledged that risk actors might exploit AI to make their operations extra environment friendly. Particularly, respondents “had been most involved with AI-powered phishing campaigns (79%) and automatic vulnerability exploitation (74%).”
SEE: Safety leaders are getting pissed off with AI-generated code.
“Safety groups should discover the most effective purposes for AI to maintain up with adversaries whereas additionally contemplating its present limitations — or threat creating extra work for themselves,” Matt Bromiley, an analyst on the SANS Institute, stated in a press launch.
The answer? AI implementations ought to endure an exterior assessment. Greater than two-thirds of these surveyed (68%) selected “exterior assessment” as the best method to establish AI security and safety points.
“Groups at the moment are extra sensible about AI’s present limitations” than they had been final yr, stated HackerOne Senior Options Architect Dane Sherrets in an electronic mail to TechRepublic. “People convey quite a lot of necessary context to each defensive and offensive safety that AI can’t replicate fairly but. Issues like hallucinations have additionally made groups hesitant to deploy the expertise in essential programs. Nevertheless, AI remains to be nice for growing productiveness and performing duties that don’t require deep context.”
Additional findings from the SANS 2024 AI Survey, launched this month, embody:
- 38% plan to undertake AI inside their safety technique sooner or later.
- 38.6% of respondents stated they’ve confronted shortcomings when utilizing AI to detect or reply to cyber threats.
- 40% cite authorized and moral implications as a problem to AI adoption.
- 41.8% of corporations have confronted pushback from workers who don’t belief AI selections, which SANS speculates is “resulting from lack of transparency.”
- 43% of organizations at present use AI inside their safety technique.
- AI expertise inside safety operations is most frequently utilized in anomaly detection programs (56.9%), malware detection (50.5%), and automatic incident response (48.9%).
- 58% of respondents stated AI programs wrestle to detect new threats or reply to outlier indicators, which SANS attributes to an absence of coaching knowledge.
- Of those that reported shortcomings with utilizing AI to detect or reply to cyber threats, 71% stated AI generated false positives.
Anthropic seeks enter from safety researchers on AI security measures
Generative AI maker Anthropic expanded its bug bounty program on HackerOne in August.
Particularly, Anthropic desires the hacker neighborhood to stress-test “the mitigations we use to forestall misuse of our fashions,” together with making an attempt to interrupt by way of the guardrails supposed to forestall AI from offering recipes for explosives or cyberattacks. Anthropic says it’s going to award as much as $15,000 to those that efficiently establish new jailbreaking assaults and can present HackerOne safety researchers with early entry to its subsequent security mitigation system.