Factbox-What we know about the rogue AI-agent security breaches

August 6 (Reuters) - Meta’s disclosure on Wednesday that one of its AI models exploited a vulnerability in a third-party service during cybersecurity testing adds to mounting evidence of the hacking risks posed by increasingly capable AI systems.

The incident followed Anthropic’s disclosure in July that its Claude models breached the systems of three companies, as well as OpenAI’s disclosure in the same month that an autonomous agent powered by its AI models compromised the infrastructure of AI startup Hugging Face. Reuters has reported that the OpenAI agent also compromised a customer of a second tech company, New York-based Modal Labs.

Here are some more details of the incidents:        

Com Date Model Organiz Duratio What occurred

pan ations n

y  breache

d

Met Inciden Meta An unna Not During a

a t did not med disclos cybersecurity

disclos identif third-p ed. evaluation run by

ed on y the arty independent tester

August model.  service Irregular, a

5, The . configuration

2026; Informa error

the tion re inadvertently gave

date of ported a Meta model

the it internet access.

testing was Mus Meta said the

inciden e Spark model then

t was 1.1. exploited a

not security

disclos vulnerability in a

ed. third-party

service. The

Information

reported it

breached an

unidentified

company’s systems

and altered its

internal

environment.

Irregular

characterized it

as an

evaluation-environ

ment issue, not a

sandbox escape or

sophisticated

cyber action.

Ope The GPT-5.6 AI The During controlled

nAI agent Sol and startup Hugging tests, an

began an Hugging Face autonomous agent

attempt unnamed Face intrusi escaped its

ing to , more and a on ran isolated

escape capable custome from environment,

its pre-rel r at July 11 accessed the

test ease New to July internet, and

environ model. York-ba 13, breached Hugging

ment sed 2026. Face to complete

around Modal its assigned goal.

July 9, Labs. The activity

2026. continued for days

  and was not

  detected by OpenAI

until after it was

contained and the

FBI was informed.

Ant The Claude All Not During

hro earlies Opus three specifi cybersecurity

pic t 4.7, organiz ed by tests, an error

inciden Claude ations Anthrop gave Claude models

t dates Mythos remain ic. internet access,

to 5, and unnamed enabling attacks

April one .  on three

2026. unnamed Anthrop companies. The

interna ic said Opus 4.7 model

l two of accessed a real

researc them company’s

h test had not credentials and

model. detecte database after

d the mistaking it for a

activit fictional target,

y another stopped

before after recognizing

Anthrop the target was

ic real.

notifie

d them;

it

continu

ed to

reach

the

third.